feat: upload tiers with placeholder sign-in, project archive/delete, no-em-dash rule

- Anonymous caps (2 MB / 500 rows, env-tunable) with sign-in hints; demo
name-only session (HMAC cookie) lifts caps; get_current_user() is the
Phase 2 swap point for managed auth
- Archive/unarchive projects (reversible, Archived sidebar group)
- Permanent delete with typed-name confirmation; cascades datasets, runs,
uploaded + result files; logged without retaining text (design par. 9)
- Fix: job executor recreates after lifespan shutdown (dev reload broke jobs)
- Style: em dashes removed across 50 files; rule in CLAUDE.md + edit hook
- Tests: 39 total (tiers, tampered cookie, archive toggle, delete cascades)

.claude/agents/docs-sync-reviewer.md

@@ -1,6 +1,6 @@
 ---
 name: docs-sync-reviewer
-description: Keeps documentation truthful after code changes — README, CHANGELOG, DECISIONS, spec status, and design-doc references. Use after completing a feature or making any architecture-relevant choice.
+description: Keeps documentation truthful after code changes - README, CHANGELOG, DECISIONS, spec status, and design-doc references. Use after completing a feature or making any architecture-relevant choice.
 tools: Read, Grep, Glob
 model: haiku
 ---
@@ -9,13 +9,13 @@ You are the documentation-sync reviewer for the CCR Platform. Docs that lie are
 no docs; a future lab member must be able to trust every README sentence.
 
 After a change, check:
-1. **README.md** — do quickstart commands still work as written? Are new features/flags mentioned?
-2. **CHANGELOG.md** — is there an entry for user-visible changes (Keep-a-Changelog style)?
-3. **DECISIONS.md** — was any architecture/infrastructure/vendor/model choice made without an
+1. **README.md** - do quickstart commands still work as written? Are new features/flags mentioned?
+2. **CHANGELOG.md** - is there an entry for user-visible changes (Keep-a-Changelog style)?
+3. **DECISIONS.md** - was any architecture/infrastructure/vendor/model choice made without an
    entry (date, decision, why, rejected alternative)? CLAUDE.md hard rule.
-4. **docs/specs/** — does the implemented behavior match the spec? Mark spec status
+4. **docs/specs/** - does the implemented behavior match the spec? Mark spec status
    (draft → implemented) and note deviations in the spec file itself.
-5. **CLAUDE.md** — are commands, layout, and phase state still accurate?
+5. **CLAUDE.md** - are commands, layout, and phase state still accurate?
 
 Return a short list of exact edits needed (file + suggested text), or "docs are in sync."
-Do not rewrite documents wholesale — smallest truthful diff.
+Do not rewrite documents wholesale - smallest truthful diff.

.claude/agents/research-validity-reviewer.md

@@ -1,6 +1,6 @@
 ---
 name: research-validity-reviewer
-description: Reviews changes for research-validity risks — hidden preprocessing changes, missing model/construct metadata, unsupported language assumptions, reverse-scoring mishandling, overconfident user-facing claims, reproducibility gaps. Use proactively after changes to the CCR engine, scoring, warnings, exports, or the model/construct registries.
+description: Reviews changes for research-validity risks - hidden preprocessing changes, missing model/construct metadata, unsupported language assumptions, reverse-scoring mishandling, overconfident user-facing claims, reproducibility gaps. Use proactively after changes to the CCR engine, scoring, warnings, exports, or the model/construct registries.
 tools: Read, Grep, Glob, Bash
 model: inherit
 ---
@@ -11,28 +11,28 @@ honesty: transparent warnings, reproducible runs, and never overstating what CCR
 
 Review the changes you are pointed at for:
 
-1. **Preprocessing drift** — any change to text normalization, deduplication, truncation, or
+1. **Preprocessing drift** - any change to text normalization, deduplication, truncation, or
    row filtering that isn't reflected in run metadata AND documented. Silent preprocessing
    changes invalidate cross-run comparisons.
-2. **Metadata completeness** — every run must record: model id + pinned revision + usage_config,
+2. **Metadata completeness** - every run must record: model id + pinned revision + usage_config,
    construct snapshot + version + item_hash, language selection + detection summary,
    preprocessing config, scoring.adjustment_strategy, output_schema_version, warnings.
-3. **Registry bypasses** — model-specific behavior (prefixes, normalization, max length,
+3. **Registry bypasses** - model-specific behavior (prefixes, normalization, max length,
    language support) hardcoded anywhere outside packages/model_registry. E5-family models
    without "query: " prefix handling is a correctness bug.
-4. **Reverse scoring** — any code that adjusts/negates reversed-item similarities beyond the
+4. **Reverse scoring** - any code that adjusts/negates reversed-item similarities beyond the
    recorded adjustment_strategy, or aggregates that silently include/exclude reversed items.
    V1 policy: raw + flags; exclude_reversed only when explicit; all-reversed must block.
-5. **Language assumptions** — English-only logic applied to multilingual paths; warnings
+5. **Language assumptions** - English-only logic applied to multilingual paths; warnings
    removed or softened; supported-language checks bypassed.
-6. **Overconfident claims** — UI/exports/docs text implying CCR measures stance, diagnoses,
+6. **Overconfident claims** - UI/exports/docs text implying CCR measures stance, diagnoses,
    or "accuracy"; missing the relatedness-vs-stance caveat where results are presented.
-7. **Reproducibility** — generated scripts that would not run offline (platform credentials,
+7. **Reproducibility** - generated scripts that would not run offline (platform credentials,
    unpinned versions, missing prefix logic); golden/expected outputs edited without approval.
 
 Return exactly three sections:
-1. **Blocking issues** — must fix before merge (cite file:line).
-2. **Non-blocking concerns** — should fix soon.
-3. **Suggested fixes** — concrete, smallest-diff suggestions.
+1. **Blocking issues** - must fix before merge (cite file:line).
+2. **Non-blocking concerns** - should fix soon.
+3. **Suggested fixes** - concrete, smallest-diff suggestions.
 
-Be specific and cite evidence. If you find nothing, say so plainly — do not invent findings.
+Be specific and cite evidence. If you find nothing, say so plainly - do not invent findings.

.claude/agents/test-parity-reviewer.md

@@ -10,7 +10,7 @@ You are the test & parity reviewer for the CCR Platform.
 Ground rules you enforce:
 - The published `ccr_wrapper` implementation is the correctness spec. Changes to similarity,
   scoring, or preprocessing must keep (or add) parity coverage within float tolerance (atol),
-  with pinned model revisions on both paths — never compare against a moving reference.
+  with pinned model revisions on both paths - never compare against a moving reference.
 - Golden evals exist to catch silent drift: output shape, column names/order, metadata fields,
   and warning behavior. `evals/expected_outputs/` is human-approved only.
 - New behavior without tests in the same change is incomplete work, not a follow-up.
@@ -26,5 +26,5 @@ When reviewing, check:
 4. Run the relevant test commands (cd backend && python -m pytest tests/ -q) and report results.
 
 Return: (1) coverage gaps with proposed test names, (2) any weakened assertions,
-(3) parity/golden risks, (4) test run results. Write new tests when asked — smallest
+(3) parity/golden risks, (4) test run results. Write new tests when asked - smallest
 useful tests, placed beside existing suites, using existing fixtures.

.claude/commands/implement-spec.md

@@ -11,5 +11,5 @@ Follow this loop strictly (CLAUDE.md workflow):
 2. Restate the contract in 3 bullets: inputs, outputs/columns, warnings/edge cases. Flag anything the spec leaves undefined BEFORE coding.
 3. Propose an implementation plan: files to change, tests to add, smallest viable diff. Wait for approval if the plan touches architecture rules.
 4. Implement in small increments. Add/update tests in the same change.
-5. Run: cd backend && python -m pytest tests/ -q — plus registry validators if registries changed.
+5. Run: cd backend && python -m pytest tests/ -q - plus registry validators if registries changed.
 6. Summarize: changed files, test results, deviations from spec (update the spec's status/notes section), CHANGELOG entry, and whether a DECISIONS.md entry is needed.

.claude/commands/review-scoring-contract.md

@@ -1,5 +1,5 @@
 ---
-description: Audit scoring/output contract — columns, adjustment_strategy, all-reversed block, metadata completeness
+description: Audit scoring/output contract - columns, adjustment_strategy, all-reversed block, metadata completeness
 ---
 
 Audit the current scoring and output contract end to end:

.claude/hooks/protect-golden.sh

@@ -8,7 +8,7 @@ file_path="$(printf '%s' "$input" | python3 -c 'import sys,json; d=json.load(sys
 
 case "$file_path" in
   *"evals/expected_outputs/"*)
-    echo "BLOCKED: $file_path is a golden expected output. Golden files change only via explicit human approval — ask Deva to update it manually or approve a regeneration script." >&2
+    echo "BLOCKED: $file_path is a golden expected output. Golden files change only via explicit human approval - ask Deva to update it manually or approve a regeneration script." >&2
     exit 2
     ;;
   *"backend/data/"*)

.claude/hooks/py-syntax-check.sh

@@ -12,4 +12,14 @@ if [[ "$file_path" == *.py && -f "$file_path" ]]; then
     exit 2
   fi
 fi
+
+# Project style rule: no em dashes in any source/text file (CLAUDE.md hard rules).
+case "$file_path" in
+  *.py|*.js|*.jsx|*.css|*.md|*.yaml|*.yml|*.sh|*.html)
+    if [[ -f "$file_path" ]] && grep -q $'\xe2\x80\x94' "$file_path"; then
+      echo "EM DASH found in $file_path. Project rule: no em dashes - use a hyphen or restructure the sentence." >&2
+      exit 2
+    fi
+    ;;
+esac
 exit 0

.claude/rules/frontend.md

@@ -7,10 +7,10 @@ paths:
 
 # Frontend rules
 
-- Current stack: React 18 + Vite, JavaScript (TypeScript arrives in Phase 2 cleanup — do not introduce it piecemeal now).
-- All server communication through frontend/src/api.js — no fetch calls inside components.
+- Current stack: React 18 + Vite, JavaScript (TypeScript arrives in Phase 2 cleanup - do not introduce it piecemeal now).
+- All server communication through frontend/src/api.js - no fetch calls inside components.
 - The UI never computes analysis values; it renders what the API returns. If a number needs computing, it belongs in the backend.
 - Warnings from the API are always rendered (amber panel); never drop or reword warning text in the UI.
 - Accessibility floor (design §19): label every input, keyboard-navigable controls, no color-only signals, charts need an accessible alternative.
 - Keep the dependency count minimal (currently react + react-dom only); adding a UI library requires a DECISIONS.md entry.
-- After frontend changes that should ship: npm run build (outputs to backend/static/) — the API serves the built SPA.
+- After frontend changes that should ship: npm run build (outputs to backend/static/) - the API serves the built SPA.

.claude/rules/python-backend.md

@@ -8,7 +8,7 @@ paths:
 
 - Python 3.10+; type hints on public functions; docstrings explain WHY, not what.
 - FastAPI: request/response shapes via Pydantic schemas; HTTPException with clear user-facing detail strings; validation errors are 400s, missing resources 404s, state conflicts 409s.
-- All embedding/normalization/similarity math stays in the engine module — API handlers orchestrate, never compute.
+- All embedding/normalization/similarity math stays in the engine module - API handlers orchestrate, never compute.
 - Every user-facing failure path returns an actionable message (what happened + what to do), never a stack trace.
 - Tests: pytest; use the deterministic fake embedder (model "fake-deterministic" / CCR_FAKE_EMBEDDINGS=1); no torch imports in the default suite; poll async jobs with a timeout helper, never sleep-and-hope.
 - Warnings are structured objects with UPPER_SNAKE codes (design §12), never bare strings.

.claude/rules/registries.md

@@ -7,9 +7,9 @@ paths:
 # Registry & construct library rules
 
 - models.yaml is the single source of model truth: id, provider_model_id, pinned revision, supported_languages OR supported_language_set, embedding_dimension, max_seq_length, usage_config (requires_prefix, prefix strings, normalize_embeddings), tiers, operational_config, user-facing warnings.
-- The `pooling` field is DESCRIPTIVE only — models load via their own sentence-transformers config; never reimplement pooling.
+- The `pooling` field is DESCRIPTIVE only - models load via their own sentence-transformers config; never reimplement pooling.
 - E5-family entries must set requires_prefix: true with "query: " on BOTH items and texts.
-- supported_language_set values (e.g., xlm_roberta_100) must resolve to real ISO codes in language_sets.py — a bare "multilingual" label is banned.
+- supported_language_set values (e.g., xlm_roberta_100) must resolve to real ISO codes in language_sets.py - a bare "multilingual" label is banned.
 - Constructs are versioned append-only YAML: editing items/wording/flags creates a NEW version with a new item_hash; never mutate an existing version in place.
 - Every construct carries: name, version, language, items (text + reverse_scored), citation, verification_status (draft|needs_verification|verified|archived). Item wordings must be flagged until verified verbatim against the original publication.
 - After any change here, run both validators before considering the change done.

.env.example

@@ -6,7 +6,7 @@
 # Row ceiling for uploads (default 100000; hosted demo uses 20000)
 # CCR_MAX_ROWS=20000
 
-# Force the deterministic fake embedder (tests/CI only — never production)
+# Force the deterministic fake embedder (tests/CI only - never production)
 # CCR_FAKE_EMBEDDINGS=1
 
 # ---- Phase 2 (not read yet; reserved names) ----

.gitignore

@@ -4,7 +4,7 @@ __pycache__/
 .venv/
 .pytest_cache/
 
-# runtime data (DB, uploaded corpora, results) — never commit research data
+# runtime data (DB, uploaded corpora, results) - never commit research data
 backend/data/
 
 # node

CHANGELOG.md

@@ -4,6 +4,21 @@ Format: [Keep a Changelog](https://keepachangelog.com/). User-visible changes on
 
 ## [Unreleased]
 
+### Added
+- Anonymous vs signed-in upload tiers: anonymous uploads capped at 2 MB / 500 rows with a
+  clear sign-in hint; signing in lifts limits. Sign-in is a labeled placeholder (name only)
+  until managed auth arrives with lab accounts; tampered sessions are treated as anonymous.
+- Project lifecycle: archive/unarchive (reversible, collapses into an Archived sidebar
+  group) and permanent delete with type-the-name confirmation. Delete cascades to datasets,
+  runs, uploaded files, and result files, and is logged without retaining any text.
+
+### Changed
+- Project-wide style rule: no em dashes in any project text (enforced by an edit hook).
+
+### Fixed
+- Job worker survives lifespan restarts (dev reload previously killed job submission
+  permanently).
+
 ### Fixed
 - Project sidebar redesigned for growing lists: always-visible search, recency groups
   (Today / This week / Earlier), per-project run count + relative last-activity time,
@@ -19,7 +34,7 @@ Format: [Keep a Changelog](https://keepachangelog.com/). User-visible changes on
   results dashboard with descriptive stats + per-item loadings + top/bottom texts,
   data-quality warnings, CSV export mirroring ccr_wrapper shape, reproducibility metadata).
 - Model registry (packages/model_registry) and versioned construct library
-  (packages/construct_library) with validators — not yet wired into the app (specs 0003/0004).
+  (packages/construct_library) with validators - not yet wired into the app (specs 0003/0004).
 - Claude Code project config: CLAUDE.md, path-scoped rules, 3 review agents, 4 commands,
   golden-file protection + Python syntax hooks.
 
@@ -30,8 +45,8 @@ Format: [Keep a Changelog](https://keepachangelog.com/). User-visible changes on
   warnings for language mismatch, uncertain detection, and model/language coverage
   (multilingual model checked against its real ISO language set). (spec 0001/0003)
 - Very-short-text warning (< 4 words) with affected-row samples. (spec 0001)
-- Downloadable Python reproduction script per run — offline-runnable, embeds construct
-  items, model revision, and E5 prefix logic — plus pinned requirements file. (spec 0002)
+- Downloadable Python reproduction script per run - offline-runnable, embeds construct
+  items, model revision, and E5 prefix logic - plus pinned requirements file. (spec 0002)
 - Model registry drives the model dropdown, validation, prefixes, and metadata:
   MiniLM (default), E5-Large-v2, Multilingual-E5-base. (spec 0003)
 - Construct library loaded from versioned YAML files with item hashes, citations, and
@@ -40,9 +55,9 @@ Format: [Keep a Changelog](https://keepachangelog.com/). User-visible changes on
   block, construct snapshot, model revision, environment pins.
 
 ### Changed
-- All warnings are structured objects (code/severity/message/count/rows) — UI renders codes.
+- All warnings are structured objects (code/severity/message/count/rows) - UI renders codes.
 - /api/models response shape changed (registry ids like "all-minilm-l6-v2" replace provider ids).
-- Delimiter detection restricted to real candidates (, ; tab |) — fixes single-column
+- Delimiter detection restricted to real candidates (, ; tab |) - fixes single-column
   sentence CSVs being split on spaces.
 
 ### Removed

CLAUDE.md

@@ -10,20 +10,20 @@ Current state: Phase 0 (visible features on the demo codebase). Roadmap: @ROADMA
 
 ## Commands
 
-- Backend tests: `cd backend && python -m pytest tests/ -q`  (fast — fake embedder, no torch needed)
+- Backend tests: `cd backend && python -m pytest tests/ -q`  (fast - fake embedder, no torch needed)
 - Run app locally: `./run.sh` → http://127.0.0.1:8000
 - Frontend build: `cd frontend && npm install && npm run build`  (outputs to backend/static/)
 - Frontend dev: `cd frontend && npm run dev`  (proxies /api to :8000)
 - Validate registries: `python packages/model_registry/validate_models.py && python packages/construct_library/validate_constructs.py`
 - Real-model smoke test: `python scripts/verify_install.py`  (downloads MiniLM once)
 
-## Architecture rules (from design §15 — non-negotiable)
+## Architecture rules (from design §15 - non-negotiable)
 
 - Analysis/NLP logic lives ONLY in the engine (today `backend/app/ccr.py`; Phase 1 extracts it to `packages/ccr_engine/`). Never in UI components or API route handlers.
 - ALL model-specific behavior (prefixes, normalization, dims, max length, language support) goes through `packages/model_registry/models.yaml`. No hardcoded model behavior anywhere.
-- Predefined constructs live ONLY in `packages/construct_library/constructs/*.yaml` — versioned, append-only; edits create new versions. (Migration from `backend/app/seed_constructs.py` is a Phase 0 task.)
+- Predefined constructs live ONLY in `packages/construct_library/constructs/*.yaml` - versioned, append-only; edits create new versions. (Migration from `backend/app/seed_constructs.py` is a Phase 0 task.)
 - Every analysis run must produce metadata + a reproduction script (design §14). Exports mirror `ccr_wrapper` output shape.
-- MiniLM (all-MiniLM-L6-v2) is the default model — it is the CCR reference model. E5-family models REQUIRE the "query: " prefix on both items and texts.
+- MiniLM (all-MiniLM-L6-v2) is the default model - it is the CCR reference model. E5-family models REQUIRE the "query: " prefix on both items and texts.
 - The published `ccr_wrapper` implementation is the correctness spec (parity tests, design §16).
 
 ## Hard rules
@@ -31,10 +31,11 @@ Current state: Phase 0 (visible features on the demo codebase). Roadmap: @ROADMA
 - Never silently change output column names (bump output_schema_version in metadata instead).
 - Never remove or weaken warnings to make tests pass.
 - Never log uploaded text content.
-- Never edit files under `evals/expected_outputs/` — golden outputs change only via explicit human approval (enforced by a PreToolUse hook).
+- Never edit files under `evals/expected_outputs/` - golden outputs change only via explicit human approval (enforced by a PreToolUse hook).
 - No new infrastructure (Redis, GPU, new vendors, new services) without a DECISIONS.md entry.
 - Reverse scoring: v1 exports raw similarities + flags only; adjustment_strategy is a recorded parameter (design §11). Do not invent adjustments.
 - Do not restructure directories big-bang; follow the strangler steps (design §7).
+- No em dashes anywhere in project text: UI strings, docs, comments, commit messages, YAML. Use a hyphen or restructure the sentence (Deva's rule, 2026-07-09; enforced by a PostToolUse hook).
 
 ## Workflow
 
@@ -45,10 +46,10 @@ Current state: Phase 0 (visible features on the demo codebase). Roadmap: @ROADMA
 
 ## PI working norms
 
-- Mohammad (2026-07-09): "you can always suggest better options that I might not know about —
+- Mohammad (2026-07-09): "you can always suggest better options that I might not know about -
   take initiative (newer NLP models, lighter options, less expensive options)." Consequence:
   when a decision involves a model, service, or cost, ALWAYS check for newer/lighter/cheaper
-  candidates and present the better option with evidence — recommending beats asking.
+  candidates and present the better option with evidence - recommending beats asking.
   New models enter via the registry + benchmark-before-adopt (golden/parity harness), never
   by swapping defaults silently. Cost options favor free tiers and annual-friendly services.
 

DECISIONS.md

@@ -1,32 +1,48 @@
 # Decisions Log
 
 > One entry per architecture/infrastructure/vendor/model decision.
-> Format: date — decision — why — rejected alternative — revisit trigger.
+> Format: date - decision - why - rejected alternative - revisit trigger.
 > CLAUDE.md hard rule: no new infrastructure without an entry here.
 
-## 2026-07-09 — Repo seeded from demo; demo stays live
+## 2026-07-09 - Repo seeded from demo; demo stays live
 Mainline moves to this repo (future home: Culture-and-Morality-Lab org). The public demo repo
-and HF Space remain deployed and untouched — the PI links people to it. Rejected: evolving the
+and HF Space remain deployed and untouched - the PI links people to it. Rejected: evolving the
 demo repo in place (would couple experiments to a live artifact). Revisit: after Phase 1, repo
 transfers to the lab org (design §7 step 7).
 
-## 2026-07-09 — Postgres jobs table over Redis/RQ (design §6)
+## 2026-07-09 - Postgres jobs table over Redis/RQ (design §6)
 Durable queue via SELECT ... FOR UPDATE SKIP LOCKED + leases + sweeper; job state already in
 the DB. Rejected: Redis/RQ (a third stateful service before pressure exists). Trigger to
 revisit: multiple workers, scheduling needs, sustained queue depth.
 
-## 2026-07-09 — Managed auth (Supabase Auth recommended), staged Google-first (design §8, §4.1)
+## 2026-07-09 - Managed auth (Supabase Auth recommended), staged Google-first (design §8, §4.1)
 Rejected: custom password auth (hashing, reset flows, verification email, lockouts = hidden
 security work). Staged: Google sign-in first (zero email infra), email/password second within
 the same provider. Trigger: PI sign-off starts Phase 2.
 
-## 2026-07-09 — MiniLM default; e5-large-v2 strong option; multilingual-E5 (design §13)
+## 2026-07-09 - MiniLM default; e5-large-v2 strong option; multilingual-E5 (design §13)
 MiniLM is the CCR reference model → comparability with the published method; E5 models require
 "query: " prefixes (encoded in registry usage_config). Rejected: e5-large-v2 as default
 (slow on CPU, breaks anonymous inline tier; cross-family scores not comparable). Trigger:
 lab validation study could change defaults.
 
-## 2026-07-09 — Reverse scoring: raw + flags only in v1 (design §11)
+## 2026-07-09 - Reverse scoring: raw + flags only in v1 (design §11)
 Aggregates exclude reversed items only when explicitly required; all-reversed blocks aggregate.
 Rejected: sign-negation by default (silently invented psychometrics). Trigger: lab methods
 decision after discussion with PI.
+
+## 2026-07-09 - Placeholder demo sign-in for tier testing (not security)
+Anonymous caps (2 MB / 500 rows, env-tunable) enforce the design §5.1 tiers now; a
+name-only signed-cookie session lifts them so the tiered UX is testable before Supabase
+auth lands. get_current_user() is the single Phase 2 integration point. Rejected: waiting
+for real auth (tiers untestable) and client-only gating (trivially bypassed).
+
+## 2026-07-09 - Open archive/delete until accounts exist
+Any visitor can archive/delete any project on the shared instance: there are no owners yet,
+and mistake-cleanup matters more than protection on a demo. Delete requires typing the
+project name, cascades to all files/rows, and is logged without text (design §9). Phase 2
+adds ownership checks to the same endpoints.
+
+## 2026-07-09 - No em dashes in project text (Deva)
+Style rule across UI, docs, comments, commits. Enforced by a PostToolUse hook; swept 50
+existing files.

DEPLOY.md

@@ -1,6 +1,6 @@
 # Deploying the demo
 
-## Option A — Hugging Face Spaces (free, recommended for the demo)
+## Option A - Hugging Face Spaces (free, recommended for the demo)
 
 Free Docker Spaces: 2 vCPU / 16 GB RAM, no credit card. Bonus: the original
 CCR online tool lives on HF Spaces, so the prototype sits where the CCR
@@ -34,19 +34,19 @@ community already works.
    First build takes ~5–10 min (model bakes into the image). Watch the
    build logs in the Space's "Logs" tab.
 
-4. Optional hardening for the public instance — in Space Settings →
+4. Optional hardening for the public instance - in Space Settings →
    Variables, set `CCR_MAX_ROWS=20000` (tighter ceiling than the
    100k default while strangers can reach it).
 
 Notes:
-- Storage is **ephemeral** — uploads/results vanish on restart or rebuild.
+- Storage is **ephemeral** - uploads/results vanish on restart or rebuild.
   Fine for a demo; the email and the in-app welcome text both say so.
 - Free Spaces sleep after ~48h without traffic. Visit the URL the
   evening before and the morning of the interview so it's warm.
 - The direct app URL (no HF frame) is
-  `https://<username>-ccr-platform.hf.space` — send that one.
+  `https://<username>-ccr-platform.hf.space` - send that one.
 
-## Option B — Google Cloud Run (few dollars, more "prod-like" URL)
+## Option B - Google Cloud Run (few dollars, more "prod-like" URL)
 
 ```bash
 gcloud run deploy ccr-platform \
@@ -62,7 +62,7 @@ gcloud run deploy ccr-platform \
 
 - `--min-instances 1`: no cold starts while he plays with it (~a few
   dollars for the week; delete the service after the process ends).
-- `--max-instances 1`: SQLite + in-process queue assume one instance —
+- `--max-instances 1`: SQLite + in-process queue assume one instance -
   documented demo trade-off, not an oversight.
 
 ## Pre-send checklist (either host)
@@ -72,9 +72,9 @@ gcloud run deploy ccr-platform \
    Satisfaction with Life → Run → results render → Export CSV downloads.
 3. Second run (Individualism) finishes in seconds (model + item cache warm).
 4. Upload your own messy CSV (Excel export with a BOM, or semicolon-
-   delimited) — parses, and any fallback is flagged in the UI.
+   delimited) - parses, and any fallback is flagged in the UI.
 5. Upload rejects a bogus file (.txt/.exe) with a clean error.
-6. Refresh the page — SPA loads, project still listed.
+6. Refresh the page - SPA loads, project still listed.
 7. Morning of the interview: open the URL once (warm the instance),
    re-run step 2 quickly.
 

Dockerfile

@@ -1,4 +1,4 @@
-# CCR Platform — Cloud Run / container deployment.
+# CCR Platform - Cloud Run / container deployment.
 # The embedding model is baked into the image so the first request
 # doesn't trigger a ~90 MB download (critical for demo cold starts).
 
@@ -15,7 +15,7 @@ WORKDIR /srv
 COPY backend/requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Pre-download ALL offered models into the image layer — a user picking a
+# Pre-download ALL offered models into the image layer - a user picking a
 # non-default model must not trigger a multi-hundred-MB download mid-job
 # (looks like a hang). Make the cache usable by any runtime UID.
 RUN python -c "from sentence_transformers import SentenceTransformer; \
@@ -28,7 +28,7 @@ RUN python -c "from sentence_transformers import SentenceTransformer; \
 COPY backend/app ./app
 COPY backend/static ./static
 
-# Demo note: SQLite + uploads live on the container's ephemeral disk —
+# Demo note: SQLite + uploads live on the container's ephemeral disk -
 # data resets on restart/redeploy. Acceptable for a demo; use
 # Postgres + S3/GCS object storage before any real use.
 EXPOSE 7860

README.md

@@ -1,4 +1,4 @@
-# CCR Platform — Culture and Morality Lab
+# CCR Platform - Culture and Morality Lab
 
 Web platform for **Contextualized Construct Representation (CCR)** psychological text analysis
 ([method](https://github.com/Ali-Omrani/CCR) · [EMNLP 2024](https://aclanthology.org/2024.emnlp-main.151/)).
@@ -20,7 +20,7 @@ reproducibility record.
 ```
 
 ```bash
-# fast test suite (deterministic fake embedder — no torch needed)
+# fast test suite (deterministic fake embedder - no torch needed)
 cd backend && pip install -r requirements-dev.txt && python -m pytest tests/ -q
 
 # validate the registries
@@ -37,7 +37,7 @@ python scripts/verify_install.py
 backend/            FastAPI app + tests (current working application)
 frontend/           React/Vite SPA source (prebuilt copy served from backend/static/)
 packages/
-  model_registry/   models.yaml — single source of model truth + validator
+  model_registry/   models.yaml - single source of model truth + validator
   construct_library/ versioned construct YAMLs + validator (append-only)
   ccr_engine/       Phase 1 target: pure analysis package (see spec 0005)
 docs/
@@ -49,7 +49,7 @@ evals/              golden datasets + expected outputs (human-approved only)
 
 ## Working on this repo
 
-Read `CLAUDE.md` first — it carries the architecture rules and hard rules (enforced partly by
+Read `CLAUDE.md` first - it carries the architecture rules and hard rules (enforced partly by
 hooks). Features start from `docs/specs/`; decisions land in `DECISIONS.md`; user-visible
 changes land in `CHANGELOG.md`. The published `ccr_wrapper` implementation is the correctness
 spec for the engine.

ROADMAP.md

@@ -1,23 +1,23 @@
-# Roadmap (from design doc §4 / §20 — capacity: one person, ≤20 hrs/week)
+# Roadmap (from design doc §4 / §20 - capacity: one person, ≤20 hrs/week)
 
-## Phase 0 — Visible progress (now)
-- [x] Descriptive stats (mean/SD/min/max) — shipped in demo codebase
+## Phase 0 - Visible progress (now)
+- [x] Descriptive stats (mean/SD/min/max) - shipped in demo codebase
 - [x] Spec 0001: short-text + language warnings (structured warning objects)
 - [x] Spec 0002: offline-runnable Python script export + pinned requirements
 - [x] Spec 0003: registry-driven models (E5 prefixes!) + language selection
 - [x] Spec 0004: construct library from versioned YAML (assistant's questionnaires land here)
 
-## Phase 1 — Engine hardening (weeks 3–5)
+## Phase 1 - Engine hardening (weeks 3–5)
 - [ ] Spec 0005: extract packages/ccr_engine + golden evals + ccr_wrapper parity tests
 - [ ] Pin model revisions in models.yaml (remove PIN_ME)
 - [ ] Registry validation in CI
 
-## Phase 2 — Persistence + auth (weeks 6–9, AFTER §22 sign-off)
+## Phase 2 - Persistence + auth (weeks 6–9, AFTER §22 sign-off)
 - [ ] Supabase Auth (Google first, then email/password) · projects · Postgres metadata + jobs
 - [ ] Object storage behind the existing storage interface · TTL cleanup (anonymous 48h)
 - [ ] Anonymous caps per design §5.1 starter table
 
-## Phase 3 — Production-lite launch
+## Phase 3 - Production-lite launch
 - [ ] Repo → Culture-and-Morality-Lab org · staging + production hosting (Option B)
 - [ ] Backups, monitoring, privacy wording (lab-provided), verified-constructs workflow
 

WORKLOG.md

@@ -1,6 +1,6 @@
 # Worklog pointer
 
 Hours and task-level logging live in the private lab worklog
-(`../CAM-Lab/worklog.md` on Deva's machine — not committed here).
+(`../CAM-Lab/worklog.md` on Deva's machine - not committed here).
 
 This file exists so repo docs can reference a stable path; keep private details out of the repo.

backend/app/auth_demo.py

@@ -0,0 +1,64 @@
+"""Placeholder sign-in for tier testing (design doc §4.1 staged rollout).
+
+This is deliberately NOT security. It exists so the anonymous-vs-signed-in
+upload tiers have real UX and enforcement before the managed auth provider
+(Supabase: Google first, then email/password) lands in Phase 2. The swap
+plan: get_current_user() is the only integration point - Phase 2 replaces
+its body with provider JWT verification and nothing else changes.
+
+Session: HMAC-signed cookie. Secret comes from CCR_SESSION_SECRET or is
+random per process (restart signs everyone out - acceptable for a demo).
+"""
+
+from __future__ import annotations
+
+import base64
+import hashlib
+import hmac
+import os
+import secrets
+
+from fastapi import Request
+
+COOKIE_NAME = "ccr_demo_session"
+_SECRET = (os.environ.get("CCR_SESSION_SECRET") or secrets.token_hex(32)).encode()
+
+# Tier limits; env-overridable so tests and the hosted demo can tune them.
+ANON_MAX_BYTES_DEFAULT = 2 * 1024 * 1024
+ANON_MAX_ROWS_DEFAULT = 500
+
+
+def anon_max_bytes() -> int:
+    return int(os.environ.get("CCR_ANON_MAX_BYTES", ANON_MAX_BYTES_DEFAULT))
+
+
+def anon_max_rows() -> int:
+    return int(os.environ.get("CCR_ANON_MAX_ROWS", ANON_MAX_ROWS_DEFAULT))
+
+
+def _sign(payload: bytes) -> str:
+    return hmac.new(_SECRET, payload, hashlib.sha256).hexdigest()
+
+
+def create_token(name: str) -> str:
+    payload = base64.urlsafe_b64encode(name.encode()).decode()
+    return f"{payload}.{_sign(payload.encode())}"
+
+
+def verify_token(token: str | None) -> str | None:
+    """Return the signed-in name, or None for anonymous/invalid tokens."""
+    if not token or "." not in token:
+        return None
+    payload, signature = token.rsplit(".", 1)
+    if not hmac.compare_digest(signature, _sign(payload.encode())):
+        return None
+    try:
+        return base64.urlsafe_b64decode(payload.encode()).decode()
+    except Exception:
+        return None
+
+
+def get_current_user(request: Request) -> dict | None:
+    """Phase 2 integration point: replace body with provider JWT verification."""
+    name = verify_token(request.cookies.get(COOKIE_NAME))
+    return {"name": name, "tier": "member"} if name else None

backend/app/ccr.py

@@ -1,4 +1,4 @@
-"""CCR engine — Contextualized Construct Representations.
+"""CCR engine - Contextualized Construct Representations.
 
 Method (Atari, Omrani et al.): embed validated questionnaire items and the
 texts to be analyzed with a contextual sentence-embedding model, then take
@@ -7,7 +7,7 @@ similarities are the text's "loadings" on the construct; their mean is the
 overall CCR score.
 
 The embedding model is injected behind a small interface so that:
-  * production uses sentence-transformers (local, pinned, reproducible —
+  * production uses sentence-transformers (local, pinned, reproducible -
     corpora never leave the machine), and
   * tests/CI use a deterministic hash-based embedder with no ML dependency.
 """
@@ -126,7 +126,7 @@ class CCRResult:
 
 
 # Item-set embeddings are tiny and constantly reused (same construct run
-# against many corpora) — cache them per (model, exact item wording).
+# against many corpora) - cache them per (model, exact item wording).
 _item_embedding_cache: dict[tuple[str, str], np.ndarray] = {}
 
 
@@ -150,7 +150,7 @@ def run_ccr(
 ) -> CCRResult:
     """Compute CCR loadings: cosine(text, item) for every text × item pair.
 
-    Prefixes come from the model registry's usage_config — E5-family models require
+    Prefixes come from the model registry's usage_config - E5-family models require
     "query: " on BOTH sides for symmetric similarity. Prefixed strings feed the
     encoder only; raw wording is what gets hashed and exported.
     """

backend/app/construct_lib.py

@@ -1,11 +1,11 @@
-"""Construct library loader — seeds the DB from packages/construct_library/constructs/.
+"""Construct library loader - seeds the DB from packages/construct_library/constructs/.
 
 Source of truth is the versioned YAML files (spec 0004, design doc §10.1). Rules:
-  * append-only: (construct_id, version) is immutable — same version with changed
+  * append-only: (construct_id, version) is immutable - same version with changed
     items is a hard error, never a silent update;
   * item_hash uses the REFERENCE implementation from validate_constructs.py (loaded
     by file path) so validator, seeder, and metadata always agree;
-  * verification_status flows to the UI — unverified wording is visibly flagged.
+  * verification_status flows to the UI - unverified wording is visibly flagged.
 
 New questionnaires from the lab land as new YAML files; `python packages/construct_library/
 validate_constructs.py` first, then restart the app (or call sync) to pick them up.

backend/app/db.py

@@ -1,4 +1,4 @@
-"""Database setup — SQLite via SQLAlchemy.
+"""Database setup - SQLite via SQLAlchemy.
 
 SQLite is a deliberate choice for this deployment size (single-node, few
 concurrent writers). The models use no SQLite-specific features, so moving

backend/app/ingest.py

@@ -1,4 +1,4 @@
-"""Corpus ingestion — tolerant of real-world research files.
+"""Corpus ingestion - tolerant of real-world research files.
 
 Researchers upload CSVs exported from Qualtrics, Excel, R, SPSS, and
 scrapers: BOMs, latin-1 encodings, semicolon/tab delimiters, ragged rows.
@@ -34,7 +34,7 @@ def load_corpus(path: str | Path) -> tuple[pd.DataFrame, dict]:
     """Parse CSV/XLSX into a DataFrame.
 
     Returns (df, parse_info) where parse_info records the format,
-    encoding, and delimiter actually used — stored with the corpus and
+    encoding, and delimiter actually used - stored with the corpus and
     echoed into every run's reproducibility metadata.
     """
     p = Path(path)
@@ -85,7 +85,7 @@ _DELIMITER_CANDIDATES = (",", ";", "\t", "|")
 def _detect_delimiter(p: Path, encoding: str) -> str:
     """Pick the candidate delimiter most consistent across the first lines.
 
-    Single-column files (no candidate present) default to ',' — a comma parse
+    Single-column files (no candidate present) default to ',' - a comma parse
     of a delimiter-free file yields one column, which is exactly right.
     """
     try:
@@ -111,7 +111,7 @@ def _validate(df: pd.DataFrame) -> pd.DataFrame:
         raise IngestError("The file parsed but contains no data rows.")
     if len(df) > max_rows():
         raise IngestError(
-            f"File has {len(df):,} rows — above this instance's "
+            f"File has {len(df):,} rows - above this instance's "
             f"{max_rows():,}-row limit. Split the corpus or run locally."
         )
     df.columns = [str(c) for c in df.columns]

backend/app/jobs.py

@@ -1,6 +1,6 @@
 """Background job runner.
 
-Jobs run on a dedicated single-worker executor — a deliberate right-sizing:
+Jobs run on a dedicated single-worker executor - a deliberate right-sizing:
 embedding is CPU-bound, so running jobs sequentially protects the instance's
 memory and keeps per-job throughput predictable, while job state lives in
 the DB (queued → running → completed/failed) so the API and UI never depend
@@ -42,15 +42,32 @@ TOP_N = 10
 SNIPPET_LEN = 220
 
 # Single worker: sequential jobs, bounded memory. See module docstring.
-_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="ccr-job")
+# Created lazily and re-creatable: a lifespan shutdown (dev reload, test
+# client closing) must not permanently kill job submission for the process.
+import threading
+
+_executor: ThreadPoolExecutor | None = None
+_executor_lock = threading.Lock()
+
+
+def _get_executor() -> ThreadPoolExecutor:
+    global _executor
+    with _executor_lock:
+        if _executor is None:
+            _executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="ccr-job")
+        return _executor
 
 
 def submit_job(job_id: str) -> None:
-    _executor.submit(_run_job_logged, job_id)
+    _get_executor().submit(_run_job_logged, job_id)
 
 
 def shutdown_executor() -> None:
-    _executor.shutdown(wait=False, cancel_futures=True)
+    global _executor
+    with _executor_lock:
+        if _executor is not None:
+            _executor.shutdown(wait=False, cancel_futures=True)
+            _executor = None
 
 
 def recover_orphaned_jobs() -> int:
@@ -129,7 +146,7 @@ def run_job(job_id: str) -> None:
             progress_cb=progress, item_prefix=item_prefix, text_prefix=text_prefix,
         )
 
-        # Structured data-quality warnings (spec 0001) — objects, never bare strings.
+        # Structured data-quality warnings (spec 0001) - objects, never bare strings.
         W = warnings_engine.warning
         warnings: list[dict] = []
         if dropped:
@@ -141,7 +158,7 @@ def run_job(job_id: str) -> None:
         if n_dupes:
             warnings.append(W(
                 "DUPLICATE_TEXTS", "warning",
-                f"{n_dupes} duplicate text(s) detected — each is scored independently; "
+                f"{n_dupes} duplicate text(s) detected - each is scored independently; "
                 "deduplicate upstream if unintended.", count=n_dupes,
             ))
         short = warnings_engine.short_text_warning(texts)

backend/app/main.py

@@ -1,4 +1,4 @@
-"""CCR Platform — FastAPI application.
+"""CCR Platform - FastAPI application.
 
 Single deployable: serves the JSON API under /api and the prebuilt React
 dashboard as static files at /. Local-first by design: corpora, embeddings,
@@ -13,12 +13,13 @@ import json
 from contextlib import asynccontextmanager
 from pathlib import Path
 
-from fastapi import Depends, FastAPI, HTTPException, UploadFile
+from fastapi import Depends, FastAPI, HTTPException, Response, UploadFile
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, JSONResponse, PlainTextResponse
 from fastapi.staticfiles import StaticFiles
 from sqlalchemy.orm import Session
 
+from . import auth_demo
 from . import jobs as jobs_module
 from . import registry
 from .ccr import FAKE_MODEL_NAME
@@ -31,10 +32,12 @@ from .schemas import (
     ConstructCreate,
     ConstructOut,
     CorpusOut,
+    DemoLogin,
     JobCreate,
     JobOut,
     ProjectCreate,
     ProjectOut,
+    ProjectPatch,
 )
 
 MAX_UPLOAD_BYTES = 25 * 1024 * 1024  # sane lab-scale ceiling; raise deliberately
@@ -133,7 +136,7 @@ def health():
 
 @app.get("/api/models")
 def list_models():
-    """Model options from the registry (spec 0003) — never hardcoded."""
+    """Model options from the registry (spec 0003) - never hardcoded."""
     return [
         {
             "id": m.id,
@@ -153,10 +156,47 @@ def list_languages():
     return SELECTABLE_LANGUAGES
 
 
+# ---------------------------------------------------- demo auth (placeholder)
+@app.get("/api/auth/me")
+def auth_me(user: dict | None = Depends(auth_demo.get_current_user)):
+    return {
+        "signed_in": user is not None,
+        "name": user["name"] if user else None,
+        "demo": True,
+        "limits": (
+            {"max_bytes": MAX_UPLOAD_BYTES, "max_rows": None}
+            if user
+            else {"max_bytes": auth_demo.anon_max_bytes(), "max_rows": auth_demo.anon_max_rows()}
+        ),
+    }
+
+
+@app.post("/api/auth/demo/login")
+def demo_login(body: DemoLogin, response: Response):
+    """Placeholder session for tier testing. Replaced by managed auth in Phase 2."""
+    name = body.name.strip()
+    if not name:
+        raise HTTPException(400, "Name cannot be empty.")
+    response.set_cookie(
+        auth_demo.COOKIE_NAME,
+        auth_demo.create_token(name),
+        httponly=True,
+        samesite="lax",
+        max_age=7 * 24 * 3600,
+    )
+    return {"signed_in": True, "name": name, "demo": True}
+
+
+@app.post("/api/auth/logout")
+def logout(response: Response):
+    response.delete_cookie(auth_demo.COOKIE_NAME)
+    return {"signed_in": False}
+
+
 # --------------------------------------------------------------- projects
 @app.get("/api/projects", response_model=list[ProjectOut])
 def list_projects(db: Session = Depends(get_db)):
-    """Projects ordered by last activity (latest run, else creation) — the
+    """Projects ordered by last activity (latest run, else creation) - the
     project a researcher wants is almost always the one they last worked on."""
     from sqlalchemy import func
 
@@ -180,6 +220,7 @@ def list_projects(db: Session = Depends(get_db)):
                 created_at=p.created_at,
                 last_activity_at=last or p.created_at,
                 n_runs=count,
+                archived=bool(p.archived),
             )
         )
     out.sort(key=lambda x: x.last_activity_at, reverse=True)
@@ -198,9 +239,65 @@ def create_project(body: ProjectCreate, db: Session = Depends(get_db)):
         created_at=project.created_at,
         last_activity_at=project.created_at,
         n_runs=0,
+        archived=False,
+    )
+
+
+@app.patch("/api/projects/{project_id}", response_model=ProjectOut)
+def patch_project(project_id: str, body: ProjectPatch, db: Session = Depends(get_db)):
+    """Archive/unarchive - reversible, no data loss. Archived projects collapse
+    into the sidebar's Archived section and keep all datasets and runs."""
+    project = _get_or_404(db, Project, project_id)
+    if body.archived is not None:
+        project.archived = bool(body.archived)
+    db.commit()
+    return ProjectOut(
+        id=project.id,
+        name=project.name,
+        description=project.description,
+        created_at=project.created_at,
+        last_activity_at=project.created_at,
+        n_runs=0,
+        archived=bool(project.archived),
     )
 
 
+@app.delete("/api/projects/{project_id}", status_code=204)
+def delete_project(project_id: str, db: Session = Depends(get_db)):
+    """Permanent delete: removes the project, its datasets, runs, uploaded
+    files, and result files. Logged without retaining any uploaded text
+    (design doc §9). Ownership checks arrive with real accounts in Phase 2;
+    until then the shared instance is deliberately open."""
+    import logging
+
+    project = _get_or_404(db, Project, project_id)
+    corpora = db.query(Corpus).filter_by(project_id=project_id).all()
+    jobs = db.query(Job).filter_by(project_id=project_id).all()
+
+    n_files = 0
+    for corpus in corpora:
+        if corpus.path and Path(corpus.path).exists():
+            Path(corpus.path).unlink(missing_ok=True)
+            n_files += 1
+    for job in jobs:
+        if job.result_path and Path(job.result_path).exists():
+            Path(job.result_path).unlink(missing_ok=True)
+            n_files += 1
+
+    for job in jobs:
+        db.delete(job)
+    for corpus in corpora:
+        db.delete(corpus)
+    db.delete(project)
+    db.commit()
+
+    logging.getLogger("ccr.projects").info(
+        "project deleted: id=%s name=%r corpora=%d runs=%d files_removed=%d",
+        project_id, project.name, len(corpora), len(jobs), n_files,
+    )
+    return Response(status_code=204)
+
+
 # ----------------------------------------------------------------- corpora
 @app.get("/api/projects/{project_id}/corpora", response_model=list[CorpusOut])
 def list_corpora(project_id: str, db: Session = Depends(get_db)):
@@ -227,7 +324,12 @@ def list_corpora(project_id: str, db: Session = Depends(get_db)):
 
 
 @app.post("/api/projects/{project_id}/corpora", response_model=CorpusOut, status_code=201)
-async def upload_corpus(project_id: str, file: UploadFile, db: Session = Depends(get_db)):
+async def upload_corpus(
+    project_id: str,
+    file: UploadFile,
+    db: Session = Depends(get_db),
+    user: dict | None = Depends(auth_demo.get_current_user),
+):
     _get_or_404(db, Project, project_id)
 
     suffix = Path(file.filename or "upload.csv").suffix.lower()
@@ -238,6 +340,15 @@ async def upload_corpus(project_id: str, file: UploadFile, db: Session = Depends
     if len(payload) > MAX_UPLOAD_BYTES:
         raise HTTPException(413, "File exceeds the 25 MB upload limit.")
 
+    # Tier gate (design §5.1): anonymous users get strict caps; signing in
+    # lifts them. Demo sign-in for now; managed auth replaces it in Phase 2.
+    if user is None and len(payload) > auth_demo.anon_max_bytes():
+        mb = auth_demo.anon_max_bytes() // (1024 * 1024)
+        raise HTTPException(
+            413,
+            f"Anonymous uploads are limited to {mb} MB. Sign in (top right) to upload larger files.",
+        )
+
     corpus = Corpus(
         project_id=project_id, filename=file.filename, path="", n_rows=0, columns_json="[]"
     )
@@ -251,6 +362,14 @@ async def upload_corpus(project_id: str, file: UploadFile, db: Session = Depends
         dest.unlink(missing_ok=True)
         raise HTTPException(400, str(exc)) from exc
 
+    if user is None and len(df) > auth_demo.anon_max_rows():
+        dest.unlink(missing_ok=True)
+        raise HTTPException(
+            400,
+            f"Anonymous uploads are limited to {auth_demo.anon_max_rows():,} rows "
+            f"(this file has {len(df):,}). Sign in (top right) to upload larger corpora.",
+        )
+
     corpus.n_rows = int(len(df))
     corpus.columns_json = json.dumps(list(df.columns))
     corpus.parse_info_json = json.dumps(parse_info)

backend/app/models.py

@@ -2,7 +2,7 @@
 
 IDs are UUID strings (portable across SQLite/Postgres). JSON-ish payloads
 (column lists, construct items, job metadata/summaries) are stored as JSON
-text — they are read-mostly blobs, not queried relationally.
+text - they are read-mostly blobs, not queried relationally.
 """
 
 import uuid
@@ -28,6 +28,7 @@ class Project(Base):
     id: Mapped[str] = mapped_column(String(32), primary_key=True, default=_uuid)
     name: Mapped[str] = mapped_column(String(200))
     description: Mapped[str] = mapped_column(Text, default="")
+    archived: Mapped[bool] = mapped_column(Boolean, default=False)
     created_at: Mapped[str] = mapped_column(String(32), default=_now)
 
 

backend/app/registry.py

@@ -1,4 +1,4 @@
-"""Model registry loader — the app-side reader of packages/model_registry/models.yaml.
+"""Model registry loader - the app-side reader of packages/model_registry/models.yaml.
 
 Single source of model truth (design doc §13): the UI dropdown, backend validation,
 prefix handling, language-support warnings, run metadata, and generated reproduction
@@ -57,7 +57,7 @@ class ModelConfig:
         return None if self.revision in ("PIN_ME", "", None) else self.revision
 
     def supports_language(self, iso_code: str) -> bool:
-        if not self.supported_languages:  # unknown coverage — never block/warn on it
+        if not self.supported_languages:  # unknown coverage - never block/warn on it
             return True
         return iso_code.lower() in self.supported_languages
 

backend/app/reproducibility.py

@@ -1,7 +1,7 @@
 """Reproduction-script generation (spec 0002, design doc §14).
 
-The generated script is built ONLY from the run's stored metadata — never from live
-state — so it reproduces what actually ran. It must be runnable outside the platform:
+The generated script is built ONLY from the run's stored metadata - never from live
+state - so it reproduces what actually ran. It must be runnable outside the platform:
 input CSV + Python + internet for the (pinned) model download. No platform credentials.
 """
 
@@ -108,7 +108,7 @@ def main(csv_path: str) -> None:
 
     work.to_csv("reproduced_results.csv", index=False)
     print(f"Wrote reproduced_results.csv ({{len(work)}} rows, {{sims.shape[1]}} items).")
-    print("Compare against the platform export — values should match to ~1e-5.")
+    print("Compare against the platform export - values should match to ~1e-5.")
 
 
 if __name__ == "__main__":

backend/app/schemas.py

@@ -15,6 +15,15 @@ class ProjectOut(BaseModel):
     created_at: str
     last_activity_at: str = ""  # latest run creation, else project creation
     n_runs: int = 0
+    archived: bool = False
+
+
+class ProjectPatch(BaseModel):
+    archived: bool | None = None
+
+
+class DemoLogin(BaseModel):
+    name: str = Field(min_length=1, max_length=80)
 
 
 class CorpusOut(BaseModel):

backend/app/warnings_engine.py

@@ -1,12 +1,12 @@
 """Structured data-quality warnings (spec 0001, design doc §12).
 
-Every warning is an object — {code, severity, message, count?, affected_rows_sample?} —
+Every warning is an object - {code, severity, message, count?, affected_rows_sample?} -
 never a bare string. Codes are UPPER_SNAKE and stable: downstream notebooks and the UI
 key off them. Severity: "info" (status, not a problem) | "warning" (proceed with care).
 Language detection is corpus-level only; short texts are exactly where detection is
 unreliable, so uncertainty is reported instead of guessed away.
 
-Deviation from spec 0001 (recorded there): langdetect instead of lingua — pure-Python,
+Deviation from spec 0001 (recorded there): langdetect instead of lingua - pure-Python,
 ~1 MB vs ~100 MB wheels; seeded for determinism. Upgrade path preserved by recording
 detector + version in metadata.
 """
@@ -75,7 +75,7 @@ def detect_corpus_language(texts: list[str], selected: str) -> tuple[LanguageRes
     except Exception:
         detector_version = "unknown"
 
-    DetectorFactory.seed = 0  # determinism — same corpus, same result, every run
+    DetectorFactory.seed = 0  # determinism - same corpus, same result, every run
 
     detectable = [t for t in texts if len(t.split()) >= DETECT_MIN_TOKENS][:DETECT_SAMPLE_MAX]
     warnings: list[dict] = []

backend/static/assets/{index-BDrO2FRr.css → index-Cu39pU_-.css}

@@ -1 +1 @@
-:root{--maroon: #7a1f3d;--maroon-dark: #5e1730;--ink: #1d2129;--muted: #667085;--line: #e5e7eb;--bg: #f7f7f8;--card: #ffffff;--ok: #157f3d;--err: #b42318;--accent-soft: #f6ebef;--control-height: 40px}*{box-sizing:border-box}body{margin:0;font-family:-apple-system,BlinkMacSystemFont,Segoe UI,Roboto,Helvetica Neue,sans-serif;color:var(--ink);background:var(--bg);font-size:14.5px;line-height:1.5}.app{display:flex;flex-direction:column;min-height:100vh}.header{background:var(--maroon);color:#fff;padding:14px 28px;display:flex;align-items:baseline;flex-wrap:wrap;gap:14px}.header h1{flex:0 0 auto;font-size:17px;margin:0;font-weight:650;letter-spacing:.2px;white-space:nowrap}.header .sub{flex:1 1 280px;min-width:0;font-size:12.5px;opacity:.85}.layout{display:flex;flex:1;min-width:0;min-height:0}.sidebar{width:250px;background:var(--card);border-right:1px solid var(--line);padding:18px 14px;flex-shrink:0;display:flex;flex-direction:column;min-height:0}.sidebar h2{font-size:11.5px;text-transform:uppercase;letter-spacing:.7px;color:var(--muted);margin:0 0 10px 4px;display:flex;align-items:center;gap:8px}.sidebar h2 .count{background:var(--bg);border:1px solid var(--line);border-radius:999px;padding:0 8px;font-size:10.5px;letter-spacing:0;color:var(--muted)}.sidebar-filter{width:100%;padding:7px 10px;margin-bottom:10px;border:1px solid var(--line);border-radius:8px;font:inherit;font-size:13px;background:#fff;color:var(--ink)}.sidebar-filter:focus{outline:none;border-color:var(--maroon)}.project-list{flex:1;min-height:0;overflow-y:auto;margin:0 -4px;padding:0 4px 4px}.group-label{font-size:10.5px;text-transform:uppercase;letter-spacing:.6px;color:var(--muted);margin:10px 4px 5px}div:first-child>.group-label{margin-top:2px}.project-item{display:block;width:100%;text-align:left;padding:9px 12px;margin-bottom:5px;border:1px solid transparent;border-radius:8px;background:none;cursor:pointer;font:inherit;color:var(--ink)}.project-item:hover{background:var(--bg)}.project-item.active{background:var(--accent-soft);border-color:var(--maroon);font-weight:600}.project-item .project-name{display:block;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}.project-item .date{display:block;font-size:11.5px;color:var(--muted);font-weight:400}.project-create{margin-top:12px;padding-top:12px;border-top:1px solid var(--line)}.project-create>button{width:100%}.main{flex:1;padding:22px 28px;overflow-y:auto;min-width:0}.card{background:var(--card);border:1px solid var(--line);border-radius:8px;padding:18px 20px;margin-bottom:16px}.card h3{margin:0 0 4px;font-size:15px}.card .hint{color:var(--muted);font-size:12.5px;margin:0 0 12px}.step-badge{display:inline-flex;align-items:center;justify-content:center;width:21px;height:21px;border-radius:50%;background:var(--maroon);color:#fff;font-size:12px;font-weight:700;margin-right:8px;vertical-align:-3px}button{white-space:nowrap}button.primary{background:var(--maroon);color:#fff;border:none;min-height:var(--control-height);display:inline-flex;align-items:center;justify-content:center;padding:0 18px;border-radius:8px;font:inherit;font-weight:600;cursor:pointer;line-height:1.2}button.primary:hover{background:var(--maroon-dark)}button.primary:disabled{background:#c9ccd1;cursor:not-allowed}button.ghost{background:none;border:1px solid var(--line);color:var(--ink);min-height:var(--control-height);display:inline-flex;align-items:center;justify-content:center;padding:0 14px;border-radius:8px;font:inherit;cursor:pointer;line-height:1.2}button.ghost:hover{border-color:var(--maroon);color:var(--maroon)}button.linkish{background:none;border:none;color:var(--maroon);font:inherit;cursor:pointer;padding:0;text-decoration:underline}input[type=text],textarea,select{width:100%;padding:8px 10px;border:1px solid var(--line);border-radius:8px;font:inherit;background:#fff;color:var(--ink)}input[type=text],select{height:var(--control-height)}button:focus-visible,input:focus-visible,textarea:focus-visible,select:focus-visible{outline:2px solid rgba(122,31,61,.38);outline-offset:2px}input[type=file]{display:block;max-width:100%;margin-top:6px;font-size:13px;color:var(--muted)}input[type=file]::file-selector-button{background:#fff;border:1px solid var(--line);color:var(--ink);padding:7px 14px;border-radius:8px;font:inherit;font-size:13px;cursor:pointer;margin-right:10px}input[type=file]::file-selector-button:hover{border-color:var(--maroon);color:var(--maroon)}.row>button{align-self:flex-end;margin-bottom:1px}textarea{resize:vertical}label.field{display:block;margin-bottom:10px;font-size:13px;font-weight:600}label.field>*{margin-top:4px;font-weight:400}.row{display:flex;gap:14px;flex-wrap:wrap}.row>*{min-width:0}.row>.grow{flex:1;min-width:min(220px,100%)}.language-control{min-width:170px}.model-control{min-width:260px}.run-settings{display:grid;grid-template-columns:minmax(160px,230px) minmax(320px,720px) max-content;justify-content:start;gap:14px;align-items:end}.run-settings .field{margin-bottom:0}.run-button{min-width:180px;height:var(--control-height)}.construct-row{display:grid;grid-template-columns:minmax(320px,1120px) max-content;justify-content:start;align-items:end;gap:14px}.results-toolbar{display:flex;align-items:center;justify-content:space-between;gap:14px;margin-bottom:14px;flex-wrap:wrap}.result-actions{justify-content:flex-end}.result-actions a{display:inline-flex;text-decoration:none}.pill{display:inline-block;padding:2px 10px;border-radius:999px;font-size:11.5px;font-weight:600}.pill.completed{background:#e6f4ea;color:var(--ok)}.pill.running{background:#fff3e0;color:#b45309}.pill.queued{background:#eef2f7;color:var(--muted)}.pill.failed{background:#fdecea;color:var(--err)}.progress-track{background:var(--line);border-radius:999px;height:7px;overflow:hidden}.progress-fill{background:var(--maroon);height:100%;transition:width .4s ease}.warnings{background:#fff8e6;border:1px solid #f2dfa8;color:#7a5b00;border-radius:8px;padding:10px 14px}.error-banner{background:#fdecea;color:var(--err);border:1px solid #f5c6c0;padding:10px 14px;border-radius:8px;margin-bottom:14px;font-size:13px}.table-wrap{width:100%;overflow-x:auto}table.docs{width:100%;border-collapse:collapse;font-size:13px}table.docs th{text-align:left;color:var(--muted);font-size:11.5px;text-transform:uppercase;letter-spacing:.5px;padding:6px 8px;border-bottom:1px solid var(--line)}table.docs td{padding:7px 8px;border-bottom:1px solid var(--bg);vertical-align:top;overflow-wrap:anywhere}table.docs td.score{font-variant-numeric:tabular-nums;font-weight:600;white-space:nowrap}.stat-grid{display:flex;gap:12px;flex-wrap:wrap;margin-bottom:4px}.stat{flex:1;min-width:110px;background:var(--bg);border-radius:8px;padding:10px 14px}.stat .v{font-size:20px;font-weight:700;font-variant-numeric:tabular-nums}.stat .k{font-size:11.5px;color:var(--muted);text-transform:uppercase;letter-spacing:.5px}.item-bar-row{display:flex;align-items:center;gap:10px;margin-bottom:7px}.item-bar-label{flex:1;font-size:12.5px;min-width:0;overflow-wrap:anywhere}.item-bar-track{flex:1.2;background:var(--bg);border-radius:999px;height:10px}.item-bar-fill{background:var(--maroon);opacity:.85;height:100%;border-radius:999px}.item-bar-val{width:52px;text-align:right;font-variant-numeric:tabular-nums;font-size:12.5px;font-weight:600}.construct-items{margin:8px 0 0;padding-left:20px;color:var(--muted);font-size:12.5px}.construct-items li{margin-bottom:2px}.meta-footer{font-size:12px;color:var(--muted);background:var(--bg);border-radius:8px;padding:10px 14px;margin-top:14px;font-variant-numeric:tabular-nums}.meta-footer code{font-size:11.5px}.muted{color:var(--muted)}.small{font-size:12.5px}.mt{margin-top:12px}@media (max-width: 820px){body{font-size:14px}.header{padding:12px 16px;align-items:flex-start;gap:2px 12px}.header h1{font-size:16px}.header .sub{flex:1 1 210px;font-size:12px;line-height:1.35}.layout{display:block}.sidebar{width:100%;border-right:0;border-bottom:1px solid var(--line);padding:14px}.project-list{max-height:220px;margin-right:0;flex:none}.project-create{border-top:0}.main{width:100%;padding:16px 14px 28px;overflow:visible}.card{padding:16px;margin-bottom:14px}.row,.run-settings,.construct-row,.results-toolbar,.result-actions{gap:10px}.row,.results-toolbar,.result-actions{flex-direction:column;align-items:stretch}.run-settings,.construct-row{grid-template-columns:1fr}.row>.grow,.construct-row>.grow,.language-control,.model-control{width:100%;min-width:0}.main .row>button,.main .row>a,.main .row>a>button,.results-toolbar>button{align-self:stretch;margin-bottom:0;width:100%}.run-button{width:100%;min-width:0}.stat-grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr))}.stat{min-width:0}.item-bar-row{display:grid;grid-template-columns:minmax(0,1fr) 56px;gap:6px 10px}.item-bar-track{grid-column:1 / -1;width:100%}.item-bar-val{width:auto}table.docs{min-width:520px}}@media (max-width: 460px){.header .sub{flex-basis:100%}.stat-grid{grid-template-columns:1fr}}
+:root{--maroon: #7a1f3d;--maroon-dark: #5e1730;--ink: #1d2129;--muted: #667085;--line: #e5e7eb;--bg: #f7f7f8;--card: #ffffff;--ok: #157f3d;--err: #b42318;--accent-soft: #f6ebef;--control-height: 40px}*{box-sizing:border-box}body{margin:0;font-family:-apple-system,BlinkMacSystemFont,Segoe UI,Roboto,Helvetica Neue,sans-serif;color:var(--ink);background:var(--bg);font-size:14.5px;line-height:1.5}.app{display:flex;flex-direction:column;min-height:100vh}.header{background:var(--maroon);color:#fff;padding:14px 28px;display:flex;align-items:baseline;flex-wrap:wrap;gap:14px}.header h1{flex:0 0 auto;font-size:17px;margin:0;font-weight:650;letter-spacing:.2px;white-space:nowrap}.header .sub{flex:1 1 280px;min-width:0;font-size:12.5px;opacity:.85}.layout{display:flex;flex:1;min-width:0;min-height:0}.sidebar{width:250px;background:var(--card);border-right:1px solid var(--line);padding:18px 14px;flex-shrink:0;display:flex;flex-direction:column;min-height:0}.sidebar h2{font-size:11.5px;text-transform:uppercase;letter-spacing:.7px;color:var(--muted);margin:0 0 10px 4px;display:flex;align-items:center;gap:8px}.sidebar h2 .count{background:var(--bg);border:1px solid var(--line);border-radius:999px;padding:0 8px;font-size:10.5px;letter-spacing:0;color:var(--muted)}.sidebar-filter{width:100%;padding:7px 10px;margin-bottom:10px;border:1px solid var(--line);border-radius:8px;font:inherit;font-size:13px;background:#fff;color:var(--ink)}.sidebar-filter:focus{outline:none;border-color:var(--maroon)}.project-list{flex:1;min-height:0;overflow-y:auto;margin:0 -4px;padding:0 4px 4px}.group-label{font-size:10.5px;text-transform:uppercase;letter-spacing:.6px;color:var(--muted);margin:10px 4px 5px}div:first-child>.group-label{margin-top:2px}.project-item{display:block;width:100%;text-align:left;padding:9px 12px;margin-bottom:5px;border:1px solid transparent;border-radius:8px;background:none;cursor:pointer;font:inherit;color:var(--ink)}.project-item:hover{background:var(--bg)}.project-item.active{background:var(--accent-soft);border-color:var(--maroon);font-weight:600}.project-item .project-name{display:block;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}.project-item .date{display:block;font-size:11.5px;color:var(--muted);font-weight:400}.project-create{margin-top:12px;padding-top:12px;border-top:1px solid var(--line)}.project-create>button{width:100%}.main{flex:1;padding:22px 28px;overflow-y:auto;min-width:0}.card{background:var(--card);border:1px solid var(--line);border-radius:8px;padding:18px 20px;margin-bottom:16px}.card h3{margin:0 0 4px;font-size:15px}.card .hint{color:var(--muted);font-size:12.5px;margin:0 0 12px}.step-badge{display:inline-flex;align-items:center;justify-content:center;width:21px;height:21px;border-radius:50%;background:var(--maroon);color:#fff;font-size:12px;font-weight:700;margin-right:8px;vertical-align:-3px}button{white-space:nowrap}button.primary{background:var(--maroon);color:#fff;border:none;min-height:var(--control-height);display:inline-flex;align-items:center;justify-content:center;padding:0 18px;border-radius:8px;font:inherit;font-weight:600;cursor:pointer;line-height:1.2}button.primary:hover{background:var(--maroon-dark)}button.primary:disabled{background:#c9ccd1;cursor:not-allowed}button.ghost{background:none;border:1px solid var(--line);color:var(--ink);min-height:var(--control-height);display:inline-flex;align-items:center;justify-content:center;padding:0 14px;border-radius:8px;font:inherit;cursor:pointer;line-height:1.2}button.ghost:hover{border-color:var(--maroon);color:var(--maroon)}button.linkish{background:none;border:none;color:var(--maroon);font:inherit;cursor:pointer;padding:0;text-decoration:underline}input[type=text],textarea,select{width:100%;padding:8px 10px;border:1px solid var(--line);border-radius:8px;font:inherit;background:#fff;color:var(--ink)}input[type=text],select{height:var(--control-height)}button:focus-visible,input:focus-visible,textarea:focus-visible,select:focus-visible{outline:2px solid rgba(122,31,61,.38);outline-offset:2px}input[type=file]{display:block;max-width:100%;margin-top:6px;font-size:13px;color:var(--muted)}input[type=file]::file-selector-button{background:#fff;border:1px solid var(--line);color:var(--ink);padding:7px 14px;border-radius:8px;font:inherit;font-size:13px;cursor:pointer;margin-right:10px}input[type=file]::file-selector-button:hover{border-color:var(--maroon);color:var(--maroon)}.row>button{align-self:flex-end;margin-bottom:1px}textarea{resize:vertical}label.field{display:block;margin-bottom:10px;font-size:13px;font-weight:600}label.field>*{margin-top:4px;font-weight:400}.row{display:flex;gap:14px;flex-wrap:wrap}.row>*{min-width:0}.row>.grow{flex:1;min-width:min(220px,100%)}.language-control{min-width:170px}.model-control{min-width:260px}.run-settings{display:grid;grid-template-columns:minmax(160px,230px) minmax(320px,720px) max-content;justify-content:start;gap:14px;align-items:end}.run-settings .field{margin-bottom:0}.run-button{min-width:180px;height:var(--control-height)}.construct-row{display:grid;grid-template-columns:minmax(320px,1120px) max-content;justify-content:start;align-items:end;gap:14px}.results-toolbar{display:flex;align-items:center;justify-content:space-between;gap:14px;margin-bottom:14px;flex-wrap:wrap}.result-actions{justify-content:flex-end}.result-actions a{display:inline-flex;text-decoration:none}.pill{display:inline-block;padding:2px 10px;border-radius:999px;font-size:11.5px;font-weight:600}.pill.completed{background:#e6f4ea;color:var(--ok)}.pill.running{background:#fff3e0;color:#b45309}.pill.queued{background:#eef2f7;color:var(--muted)}.pill.failed{background:#fdecea;color:var(--err)}.progress-track{background:var(--line);border-radius:999px;height:7px;overflow:hidden}.progress-fill{background:var(--maroon);height:100%;transition:width .4s ease}.warnings{background:#fff8e6;border:1px solid #f2dfa8;color:#7a5b00;border-radius:8px;padding:10px 14px}.error-banner{background:#fdecea;color:var(--err);border:1px solid #f5c6c0;padding:10px 14px;border-radius:8px;margin-bottom:14px;font-size:13px}.table-wrap{width:100%;overflow-x:auto}table.docs{width:100%;border-collapse:collapse;font-size:13px}table.docs th{text-align:left;color:var(--muted);font-size:11.5px;text-transform:uppercase;letter-spacing:.5px;padding:6px 8px;border-bottom:1px solid var(--line)}table.docs td{padding:7px 8px;border-bottom:1px solid var(--bg);vertical-align:top;overflow-wrap:anywhere}table.docs td.score{font-variant-numeric:tabular-nums;font-weight:600;white-space:nowrap}.stat-grid{display:flex;gap:12px;flex-wrap:wrap;margin-bottom:4px}.stat{flex:1;min-width:110px;background:var(--bg);border-radius:8px;padding:10px 14px}.stat .v{font-size:20px;font-weight:700;font-variant-numeric:tabular-nums}.stat .k{font-size:11.5px;color:var(--muted);text-transform:uppercase;letter-spacing:.5px}.item-bar-row{display:flex;align-items:center;gap:10px;margin-bottom:7px}.item-bar-label{flex:1;font-size:12.5px;min-width:0;overflow-wrap:anywhere}.item-bar-track{flex:1.2;background:var(--bg);border-radius:999px;height:10px}.item-bar-fill{background:var(--maroon);opacity:.85;height:100%;border-radius:999px}.item-bar-val{width:52px;text-align:right;font-variant-numeric:tabular-nums;font-size:12.5px;font-weight:600}.construct-items{margin:8px 0 0;padding-left:20px;color:var(--muted);font-size:12.5px}.construct-items li{margin-bottom:2px}.meta-footer{font-size:12px;color:var(--muted);background:var(--bg);border-radius:8px;padding:10px 14px;margin-top:14px;font-variant-numeric:tabular-nums}.meta-footer code{font-size:11.5px}.muted{color:var(--muted)}.small{font-size:12.5px}.mt{margin-top:12px}.header-auth{margin-left:auto;display:flex;align-items:center;gap:10px;color:#fff}.header-btn{background:#ffffff1f;color:#fff;border:1px solid rgba(255,255,255,.45);padding:5px 14px;border-radius:7px;font:inherit;font-size:13px;cursor:pointer}.header-btn:hover{background:#ffffff38}.project-header{display:flex;align-items:center;justify-content:space-between;gap:12px;margin-bottom:14px;flex-wrap:wrap}.project-title{font-size:17px;font-weight:650;margin-right:10px}button.danger{color:var(--err);border-color:#f0c4be}button.danger:hover{color:var(--err);border-color:var(--err)}button.danger-solid{background:var(--err)}button.danger-solid:hover{background:#93261b}button.danger-solid:disabled{background:#c9ccd1}.modal-backdrop{position:fixed;top:0;right:0;bottom:0;left:0;background:#14161a73;display:flex;align-items:center;justify-content:center;z-index:40;padding:16px}.modal{background:var(--card);border-radius:10px;padding:22px 24px;width:100%;max-width:430px;box-shadow:0 12px 40px #0000002e}.modal h3{margin:0 0 6px}@media (max-width: 820px){body{font-size:14px}.header{padding:12px 16px;align-items:flex-start;gap:2px 12px}.header h1{font-size:16px}.header .sub{flex:1 1 210px;font-size:12px;line-height:1.35}.layout{display:block}.sidebar{width:100%;border-right:0;border-bottom:1px solid var(--line);padding:14px}.project-list{max-height:220px;margin-right:0;flex:none}.project-create{border-top:0}.main{width:100%;padding:16px 14px 28px;overflow:visible}.card{padding:16px;margin-bottom:14px}.row,.run-settings,.construct-row,.results-toolbar,.result-actions{gap:10px}.row,.results-toolbar,.result-actions{flex-direction:column;align-items:stretch}.run-settings,.construct-row{grid-template-columns:1fr}.row>.grow,.construct-row>.grow,.language-control,.model-control{width:100%;min-width:0}.main .row>button,.main .row>a,.main .row>a>button,.results-toolbar>button{align-self:stretch;margin-bottom:0;width:100%}.run-button{width:100%;min-width:0}.stat-grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr))}.stat{min-width:0}.item-bar-row{display:grid;grid-template-columns:minmax(0,1fr) 56px;gap:6px 10px}.item-bar-track{grid-column:1 / -1;width:100%}.item-bar-val{width:auto}table.docs{min-width:520px}}@media (max-width: 460px){.header .sub{flex-basis:100%}.stat-grid{grid-template-columns:1fr}}

backend/static/index.html

@@ -4,8 +4,8 @@
     <meta charset="UTF-8" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <title>CCR Platform — Contextualized Construct Representations</title>
-    <script type="module" crossorigin src="/assets/index-CHG17tHs.js"></script>
-    <link rel="stylesheet" crossorigin href="/assets/index-BDrO2FRr.css">
+    <script type="module" crossorigin src="/assets/index-DOOkGzjl.js"></script>
+    <link rel="stylesheet" crossorigin href="/assets/index-Cu39pU_-.css">
   </head>
   <body>
     <div id="root"></div>

backend/tests/test_api.py

@@ -1,8 +1,8 @@
 """End-to-end API tests: project -> upload -> job -> results -> export.
 
 Run against the fake embedding backend (model_name = 'fake-deterministic'),
-exercising the full pipeline — tolerant upload parsing, the real job queue
-(worker thread + polling), results summary, warnings, and export shape —
+exercising the full pipeline - tolerant upload parsing, the real job queue
+(worker thread + polling), results summary, warnings, and export shape -
 without ML dependencies.
 """
 
@@ -81,7 +81,7 @@ def test_health(client):
 def test_seed_constructs_present(client):
     names = {c["name"] for c in client.get("/api/constructs").json()}
     assert "Satisfaction with Life" in names
-    assert "Moral Foundations — Care" in names
+    assert "Moral Foundations - Care" in names
 
 
 def test_corpus_upload_parses_columns(flow):

backend/tests/test_auth_tiers_and_lifecycle.py

@@ -0,0 +1,134 @@
+"""Demo sign-in tiers (anonymous caps lift on sign-in) and project lifecycle
+(archive is reversible; delete cascades to files and rows)."""
+
+import io
+import time
+from pathlib import Path
+
+import pytest
+from fastapi.testclient import TestClient
+
+from app.main import app
+
+
+def upload(client, project_id, name, payload: bytes):
+    return client.post(
+        f"/api/projects/{project_id}/corpora",
+        files={"file": (name, io.BytesIO(payload), "text/csv")},
+    )
+
+
+def wait_for_job(client, job_id, timeout=10.0):
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        job = client.get(f"/api/jobs/{job_id}").json()
+        if job["status"] in ("completed", "failed"):
+            return job
+        time.sleep(0.05)
+    raise TimeoutError(job_id)
+
+
+@pytest.fixture()
+def client():
+    with TestClient(app) as c:
+        yield c
+
+
+def csv_rows(n: int) -> bytes:
+    return ("text\n" + "\n".join(f"sample sentence number {i} here" for i in range(n))).encode()
+
+
+# ------------------------------------------------------------------- tiers
+def test_anonymous_row_cap_and_signin_lifts_it(client, monkeypatch):
+    monkeypatch.setenv("CCR_ANON_MAX_ROWS", "5")
+    project = client.post("/api/projects", json={"name": "Tiers"}).json()
+
+    me = client.get("/api/auth/me").json()
+    assert me["signed_in"] is False and me["limits"]["max_rows"] == 5
+
+    resp = upload(client, project["id"], "big.csv", csv_rows(10))
+    assert resp.status_code == 400
+    assert "Sign in" in resp.json()["detail"]
+
+    assert client.post("/api/auth/demo/login", json={"name": "Deva"}).status_code == 200
+    me = client.get("/api/auth/me").json()
+    assert me["signed_in"] is True and me["name"] == "Deva"
+
+    resp = upload(client, project["id"], "big.csv", csv_rows(10))
+    assert resp.status_code == 201, resp.json()
+
+    client.post("/api/auth/logout")
+    assert client.get("/api/auth/me").json()["signed_in"] is False
+
+
+def test_anonymous_size_cap(client, monkeypatch):
+    monkeypatch.setenv("CCR_ANON_MAX_BYTES", "200")
+    project = client.post("/api/projects", json={"name": "SizeCap"}).json()
+    resp = upload(client, project["id"], "big.csv", csv_rows(50))
+    assert resp.status_code == 413
+    assert "Sign in" in resp.json()["detail"]
+
+
+def test_tampered_session_cookie_is_anonymous(client):
+    client.cookies.set("ccr_demo_session", "aGFja2Vy.badsignature")
+    assert client.get("/api/auth/me").json()["signed_in"] is False
+
+
+# --------------------------------------------------------------- lifecycle
+def test_archive_toggle_is_reversible(client):
+    project = client.post("/api/projects", json={"name": "Archivable"}).json()
+    assert project["archived"] is False
+
+    patched = client.patch(f"/api/projects/{project['id']}", json={"archived": True}).json()
+    assert patched["archived"] is True
+    listed = next(p for p in client.get("/api/projects").json() if p["id"] == project["id"])
+    assert listed["archived"] is True
+
+    patched = client.patch(f"/api/projects/{project['id']}", json={"archived": False}).json()
+    assert patched["archived"] is False
+
+
+def test_delete_cascades_rows_and_files(client):
+    project = client.post("/api/projects", json={"name": "Doomed"}).json()
+    corpus = upload(client, project["id"], "corpus.csv", csv_rows(6)).json()
+
+    constructs = client.get("/api/constructs").json()
+    swls = next(c for c in constructs if c["name"] == "Satisfaction with Life")
+    job = client.post(
+        "/api/jobs",
+        json={
+            "project_id": project["id"],
+            "corpus_id": corpus["id"],
+            "construct_id": swls["id"],
+            "text_column": "text",
+            "model_name": "fake-deterministic",
+        },
+    ).json()
+    job = wait_for_job(client, job["id"])
+    assert job["status"] == "completed"
+
+    # capture file paths before deletion
+    results = client.get(f"/api/jobs/{job['id']}/results")
+    assert results.status_code == 200
+
+    resp = client.delete(f"/api/projects/{project['id']}")
+    assert resp.status_code == 204
+
+    assert client.get(f"/api/jobs/{job['id']}").status_code == 404
+    assert all(p["id"] != project["id"] for p in client.get("/api/projects").json())
+    # corpora listing for the deleted project 404s
+    assert client.get(f"/api/projects/{project['id']}/corpora").status_code == 404
+
+
+def test_delete_removes_files_on_disk(client, tmp_path):
+    import os
+
+    data_dir = Path(os.environ["CCR_DATA_DIR"])
+    project = client.post("/api/projects", json={"name": "FileCheck"}).json()
+    before = set((data_dir / "corpora").glob("*"))
+    upload(client, project["id"], "corpus.csv", csv_rows(6))
+    created = set((data_dir / "corpora").glob("*")) - before
+    assert len(created) == 1
+
+    client.delete(f"/api/projects/{project['id']}")
+    assert not created.pop().exists()

backend/tests/test_ccr.py

@@ -1,4 +1,4 @@
-"""Unit tests for the CCR engine (deterministic fake backend — no torch)."""
+"""Unit tests for the CCR engine (deterministic fake backend - no torch)."""
 
 import numpy as np
 import pytest

docs/specs/0001-data-quality-warnings.md

@@ -1,13 +1,13 @@
-# Spec 0001 — Data-quality warnings: very short texts + likely wrong language
+# Spec 0001 - Data-quality warnings: very short texts + likely wrong language
 
 **Status:** implemented (2026-07-09)
 **Phase:** 0    **Design doc ref:** §12 (language), §11 pipeline step 3 (Text QA)
-**Requested by:** Mohammad (Slack brief: "proper warnings — wrong language, super short texts, super long texts")
+**Requested by:** Mohammad (Slack brief: "proper warnings - wrong language, super short texts, super long texts")
 
 ## Problem
 Researchers upload corpora with rows that embed unreliably (very short texts) or that don't
 match the language they selected. Today the platform warns about empties, duplicates, and
-truncation-length texts, but not about short texts or language mismatch — so silent validity
+truncation-length texts, but not about short texts or language mismatch - so silent validity
 problems reach published results.
 
 ## Contract
@@ -25,7 +25,7 @@ problems reach published results.
 
 **Language detection design:**
 - Corpus-level only: sample up to 200 rows with ≥ 5 tokens; if < 20 such rows → LANGUAGE_UNCERTAIN.
-- Library: `lingua-language-detector` (pure-Python wheels, no torch) — add to requirements; record library+version in metadata.
+- Library: `lingua-language-detector` (pure-Python wheels, no torch) - add to requirements; record library+version in metadata.
 - Never block on language; user may proceed (design §12 user-override rule).
 
 **Warning object schema (all warnings migrate to this shape):**
@@ -51,12 +51,12 @@ splitting (future spec).
 - golden: sample_corpus.csv against SWLS produces byte-stable warning set
 
 ## Implementation notes
-- Warning construction centralizes in a new `warnings.py` module (engine-side, pure) — first
+- Warning construction centralizes in a new `warnings.py` module (engine-side, pure) - first
   extraction step toward packages/ccr_engine.
 - Detection runs once at job start on the parsed dataframe; result flows into metadata + warnings.
-- UI: no new components — existing warnings panel renders the message strings.
+- UI: no new components - existing warnings panel renders the message strings.
 
 ## Deviations (filled after implementation)
-- Detector: `langdetect` (seeded, deterministic) instead of lingua — ~1 MB pure-Python vs ~100 MB wheels; detector + version recorded in metadata so an upgrade stays traceable.
+- Detector: `langdetect` (seeded, deterministic) instead of lingua - ~1 MB pure-Python vs ~100 MB wheels; detector + version recorded in metadata so an upgrade stays traceable.
 - LANGUAGE_MISMATCH threshold implemented as majority-share >= 0.70 across sampled rows.
 - MODEL_NOTE (info) added: registry user_warnings surface per run (not in original spec).

docs/specs/0002-python-script-export.md

@@ -1,4 +1,4 @@
-# Spec 0002 — Reproduction script export (offline-runnable)
+# Spec 0002 - Reproduction script export (offline-runnable)
 
 **Status:** implemented (2026-07-09)
 **Phase:** 0    **Design doc ref:** §14

docs/specs/0003-model-registry-and-language-selection.md

@@ -1,4 +1,4 @@
-# Spec 0003 — Registry-driven models + language selection (brief)
+# Spec 0003 - Registry-driven models + language selection (brief)
 
 **Status:** implemented (2026-07-09)
 **Phase:** 0    **Design doc ref:** §12, §13
@@ -12,5 +12,5 @@ into spec-0001 warnings. Add e5-large-v2 and multilingual-e5-base to the UI via
 Validator (`validate_models.py`) runs in tests. Pin revisions before Phase 1 exit.
 
 **Key tests:** loader validation errors, prefix application for E5 (assert embedded strings
-receive "query: " — testable via fake backend capturing inputs), /api/models shape, metadata
+receive "query: " - testable via fake backend capturing inputs), /api/models shape, metadata
 records registry snapshot, MiniLM stays default.

docs/specs/0004-construct-library-yaml.md

@@ -1,4 +1,4 @@
-# Spec 0004 — Construct library from YAML (brief)
+# Spec 0004 - Construct library from YAML (brief)
 
 **Status:** implemented (2026-07-09)
 **Phase:** 0/1    **Design doc ref:** §10.1, §11
@@ -7,7 +7,7 @@
 Replace `backend/app/seed_constructs.py` seeding with `packages/construct_library/constructs/*.yaml`
 as the source of truth: loader validates via the same rules as `validate_constructs.py`,
 computes item_hash with the reference algorithm, seeds/updates DB rows keyed
-(construct_id, version) — append-only, never mutate an existing version. Run metadata embeds
+(construct_id, version) - append-only, never mutate an existing version. Run metadata embeds
 the full construct snapshot. UI shows verification_status (verified prominently; others with
 a "wording not yet verified" chip). New questionnaires from the lab = new YAML files + validator
 run; item wordings marked needs_verification until checked verbatim against the publication.

docs/specs/0005-engine-extraction-and-parity.md

@@ -1,6 +1,6 @@
-# Spec 0005 — ccr_engine extraction + reference parity tests (brief)
+# Spec 0005 - ccr_engine extraction + reference parity tests (brief)
 
-**Status:** draft (Phase 1 opener — refine after Phase 0 ships)
+**Status:** draft (Phase 1 opener - refine after Phase 0 ships)
 **Phase:** 1    **Design doc ref:** §7 (strangler steps 1–3), §10, §16
 
 Extract the pure analysis package per the strangler plan: move similarity/scoring/warnings/
@@ -8,7 +8,7 @@ stats/reproducibility logic from backend/app into `packages/ccr_engine/` with th
 `run_ccr_analysis(...)` interface (design §10); backend becomes an orchestration layer. Freeze
 behavior first with golden datasets (evals/golden_datasets → evals/expected_outputs, human-
 approved). Then add reference parity: same corpus + items through the published `ccr_wrapper`
-(pyccr) and through ccr_engine — similarities match within atol (1e-5 target), pinned model
+(pyccr) and through ccr_engine - similarities match within atol (1e-5 target), pinned model
 revisions on BOTH paths, run in the same container as production. Document any expected
 differences. Editable-install workspace wiring so backend imports the package (no sys.path hacks).
 

docs/specs/TEMPLATE.md

@@ -1,4 +1,4 @@
-# Spec NNNN — <feature name>
+# Spec NNNN - <feature name>
 
 **Status:** draft | approved | implemented (+ deviations noted below)
 **Phase:** 0 | 1 | 2 | 3    **Design doc ref:** §<n>

evals/README.md

@@ -1,10 +1,10 @@
 # Golden evals
 
-Purpose: catch silent research-behavior drift — output shape, column names/order, metadata
-fields, warning codes — per design doc §16.
+Purpose: catch silent research-behavior drift - output shape, column names/order, metadata
+fields, warning codes - per design doc §16.
 
-- `golden_datasets/` — frozen input corpora (never edit rows in place; add new files).
-- `expected_outputs/` — human-approved expected results. **Protected by a PreToolUse hook:**
+- `golden_datasets/` - frozen input corpora (never edit rows in place; add new files).
+- `expected_outputs/` - human-approved expected results. **Protected by a PreToolUse hook:**
   Claude Code cannot edit files here. To update after an approved behavior change, regenerate
   manually and review the diff yourself before committing.
 - Runner arrives with spec 0005 (engine extraction freezes behavior first). Until then, the

frontend/index.html

@@ -3,7 +3,7 @@
   <head>
     <meta charset="UTF-8" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
-    <title>CCR Platform — Contextualized Construct Representations</title>
+    <title>CCR Platform - Contextualized Construct Representations</title>
   </head>
   <body>
     <div id="root"></div>

frontend/src/App.jsx

@@ -16,12 +16,17 @@ function relativeTime(iso) {
 }
 
 function groupProjects(projects) {
-  // Buckets by last activity: Today / This week / Earlier. Projects arrive
+  // Buckets by last activity: Today / This week / Earlier, with archived
+  // projects collapsed into their own group at the bottom. Projects arrive
   // sorted by last activity (backend), so group order falls out naturally.
   const now = Date.now();
   const DAY = 86400000;
-  const groups = { Today: [], "This week": [], Earlier: [] };
+  const groups = { Today: [], "This week": [], Earlier: [], Archived: [] };
   for (const p of projects) {
+    if (p.archived) {
+      groups.Archived.push(p);
+      continue;
+    }
     const iso = p.last_activity_at || p.created_at;
     const t = new Date(iso.endsWith("Z") || iso.includes("+") ? iso : iso + "Z").getTime();
     const age = now - t;
@@ -39,14 +44,41 @@ export default function App() {
   const [newName, setNewName] = useState("");
   const [filter, setFilter] = useState("");
   const [error, setError] = useState("");
+  const [auth, setAuth] = useState(null);
+  const [showLogin, setShowLogin] = useState(false);
+  const [loginName, setLoginName] = useState("");
 
   const loadProjects = () =>
     api.listProjects().then(setProjects).catch((e) => setError(e.message));
+  const loadAuth = () => api.authMe().then(setAuth).catch(() => {});
 
   useEffect(() => {
     loadProjects();
+    loadAuth();
   }, []);
 
+  async function handleLogin(e) {
+    e.preventDefault();
+    if (!loginName.trim()) return;
+    try {
+      await api.demoLogin(loginName.trim());
+      setShowLogin(false);
+      setLoginName("");
+      await loadAuth();
+    } catch (err) {
+      setError(err.message);
+    }
+  }
+
+  async function handleLogout() {
+    try {
+      await api.logout();
+      await loadAuth();
+    } catch (err) {
+      setError(err.message);
+    }
+  }
+
   useEffect(() => {
     if (projects.length === 0) {
       setSelectedId(null);
@@ -84,8 +116,64 @@ export default function App() {
         <span className="sub">
           Contextualized Construct Representations · theory-driven psychological text analysis
         </span>
+        <span className="header-auth">
+          {auth?.signed_in ? (
+            <>
+              <span className="small">Hi, {auth.name}</span>
+              <button className="header-btn" onClick={handleLogout}>
+                Sign out
+              </button>
+            </>
+          ) : (
+            <button className="header-btn" onClick={() => setShowLogin(true)}>
+              Sign in
+            </button>
+          )}
+        </span>
       </header>
 
+      {showLogin && (
+        <div className="modal-backdrop" onClick={() => setShowLogin(false)}>
+          <div className="modal" onClick={(e) => e.stopPropagation()}>
+            <h3>Sign in</h3>
+            <p className="hint">
+              Signing in lifts the anonymous upload limits
+              {auth?.limits?.max_rows
+                ? ` (currently ${Math.round(auth.limits.max_bytes / 1048576)} MB / ${auth.limits.max_rows.toLocaleString()} rows per file)`
+                : ""}
+              .
+            </p>
+            <button className="ghost" disabled title="Arrives with lab accounts">
+              Sign in with Google (coming soon)
+            </button>
+            <form onSubmit={handleLogin} className="mt">
+              <label className="field">
+                Your name (placeholder sign-in for now)
+                <input
+                  type="text"
+                  autoFocus
+                  value={loginName}
+                  onChange={(e) => setLoginName(e.target.value)}
+                  placeholder="e.g. Mohammad"
+                />
+              </label>
+              <div className="row">
+                <button className="primary" type="submit" disabled={!loginName.trim()}>
+                  Continue
+                </button>
+                <button className="ghost" type="button" onClick={() => setShowLogin(false)}>
+                  Cancel
+                </button>
+              </div>
+            </form>
+            <p className="small muted mt">
+              This is a temporary placeholder so larger uploads can be tested. Real
+              sign-in (Google + university account) arrives with lab accounts.
+            </p>
+          </div>
+        </div>
+      )}
+
       <div className="layout">
         <aside className="sidebar">
           <h2>
@@ -158,7 +246,16 @@ export default function App() {
             </div>
           )}
           {selected ? (
-            <Workspace key={selected.id} project={selected} />
+            <Workspace
+              key={selected.id}
+              project={selected}
+              auth={auth}
+              onProjectChanged={loadProjects}
+              onProjectDeleted={() => {
+                setSelectedId(null);
+                loadProjects();
+              }}
+            />
           ) : (
             <div className="card">
               <h3>Welcome</h3>
@@ -168,7 +265,7 @@ export default function App() {
                 score distributions, and a reproducibility record for every run.
               </p>
               <p className="small muted">
-                Self-contained by design: embeddings run on this server itself — no
+                Self-contained by design: embeddings run on this server itself - no
                 third-party AI APIs. Demo instance: storage is ephemeral and may reset;
                 please don&apos;t upload sensitive or identifiable data.
               </p>

frontend/src/ResultsView.jsx

@@ -73,7 +73,7 @@ export default function ResultsView({ jobId, onBack }) {
                 <li key={i}>
                   {typeof w === "string" ? w : (
                     <>
-                      <code style={{ fontSize: 11 }}>{w.code}</code> — {w.message}
+                      <code style={{ fontSize: 11 }}>{w.code}</code> - {w.message}
                     </>
                   )}
                 </li>
@@ -91,7 +91,7 @@ export default function ResultsView({ jobId, onBack }) {
       <div className="card">
         <h3>Per-item mean loadings</h3>
         <p className="hint">
-          Mean similarity of the corpus to each scale item — a face-validity check on which
+          Mean similarity of the corpus to each scale item - a face-validity check on which
           items drive the construct signal.
         </p>
         {summary.item_means.map((m, i) => (
@@ -122,7 +122,7 @@ export default function ResultsView({ jobId, onBack }) {
       </div>
 
       <div className="meta-footer">
-        <strong>Reproducibility record</strong> — model: <code>{metadata.model}</code> (dim{" "}
+        <strong>Reproducibility record</strong> - model: <code>{metadata.model}</code> (dim{" "}
         {metadata.embedding_dim}) · items hash: <code>{metadata.items_sha256_16}</code> ·
         text column: <code>{metadata.text_column}</code> · run:{" "}
         {metadata.started_at} → {metadata.finished_at} ({metadata.duration_seconds}s) ·
@@ -130,7 +130,7 @@ export default function ResultsView({ jobId, onBack }) {
         {metadata.sentence_transformers &&
           ` · sentence-transformers ${metadata.sentence_transformers}`}
         <div className="mt small">
-          Construct reference: {metadata.construct_reference || "—"}
+          Construct reference: {metadata.construct_reference || "-"}
         </div>
       </div>
     </>

frontend/src/Workspace.jsx

@@ -2,7 +2,7 @@ import { useCallback, useEffect, useRef, useState } from "react";
 import { api } from "./api.js";
 import ResultsView from "./ResultsView.jsx";
 
-export default function Workspace({ project }) {
+export default function Workspace({ project, auth, onProjectChanged, onProjectDeleted }) {
   const [corpora, setCorpora] = useState([]);
   const [constructs, setConstructs] = useState([]);
   const [models, setModels] = useState([]);
@@ -19,8 +19,29 @@ export default function Workspace({ project }) {
   const [error, setError] = useState("");
   const [showNewConstruct, setShowNewConstruct] = useState(false);
   const [viewJobId, setViewJobId] = useState(null);
+  const [confirmDelete, setConfirmDelete] = useState(false);
+  const [deleteText, setDeleteText] = useState("");
   const fileRef = useRef(null);
 
+  async function toggleArchive() {
+    try {
+      await api.patchProject(project.id, { archived: !project.archived });
+      onProjectChanged?.();
+    } catch (err) {
+      setError(err.message);
+    }
+  }
+
+  async function handleDelete() {
+    try {
+      await api.deleteProject(project.id);
+      setConfirmDelete(false);
+      onProjectDeleted?.();
+    } catch (err) {
+      setError(err.message);
+    }
+  }
+
   const refreshJobs = useCallback(
     () => api.listJobs(project.id).then(setJobs).catch(() => {}),
     [project.id]
@@ -113,13 +134,78 @@ export default function Workspace({ project }) {
         </div>
       )}
 
-      {/* Step 1 — corpus */}
+      {/* Project header + actions */}
+      <div className="project-header">
+        <div>
+          <span className="project-title">{project.name}</span>
+          {project.archived && <span className="pill queued">archived</span>}
+        </div>
+        <div className="row">
+          <button className="ghost" onClick={toggleArchive}>
+            {project.archived ? "Unarchive" : "Archive"}
+          </button>
+          <button className="ghost danger" onClick={() => setConfirmDelete(true)}>
+            Delete
+          </button>
+        </div>
+      </div>
+
+      {confirmDelete && (
+        <div className="modal-backdrop" onClick={() => setConfirmDelete(false)}>
+          <div className="modal" onClick={(e) => e.stopPropagation()}>
+            <h3>Delete "{project.name}"?</h3>
+            <p className="hint">
+              This permanently deletes {corpora.length} dataset{corpora.length === 1 ? "" : "s"},{" "}
+              {jobs.length} run{jobs.length === 1 ? "" : "s"}, and all uploaded and result files.
+              This cannot be undone. If you might need it later, use Archive instead.
+            </p>
+            <label className="field">
+              Type the project name to confirm
+              <input
+                type="text"
+                autoFocus
+                value={deleteText}
+                onChange={(e) => setDeleteText(e.target.value)}
+                placeholder={project.name}
+              />
+            </label>
+            <div className="row">
+              <button
+                className="primary danger-solid"
+                disabled={deleteText !== project.name}
+                onClick={handleDelete}
+              >
+                Delete permanently
+              </button>
+              <button
+                className="ghost"
+                onClick={() => {
+                  setConfirmDelete(false);
+                  setDeleteText("");
+                }}
+              >
+                Cancel
+              </button>
+            </div>
+          </div>
+        </div>
+      )}
+
+      {/* Step 1 - corpus */}
       <div className="card">
         <h3>
           <span className="step-badge">1</span>Corpus
         </h3>
         <p className="hint">
           Upload a CSV or XLSX file, then choose the column containing the text to analyze.
+          {auth && !auth.signed_in && auth.limits?.max_rows && (
+            <>
+              {" "}
+              Anonymous limit: {Math.round(auth.limits.max_bytes / 1048576)} MB /{" "}
+              {auth.limits.max_rows.toLocaleString()} rows per file; sign in (top right) for
+              larger uploads.
+            </>
+          )}
         </p>
         <div className="row">
           <div className="grow">
@@ -139,7 +225,7 @@ export default function Workspace({ project }) {
             <label className="field">
               Corpus
               <select value={corpusId} onChange={(e) => setCorpusId(e.target.value)}>
-                <option value="">— select —</option>
+                <option value="">- select -</option>
                 {corpora.map((c) => (
                   <option key={c.id} value={c.id}>
                     {c.filename} ({c.n_rows.toLocaleString()} rows)
@@ -156,7 +242,7 @@ export default function Workspace({ project }) {
                 onChange={(e) => setTextColumn(e.target.value)}
                 disabled={!corpus}
               >
-                <option value="">— select —</option>
+                <option value="">- select -</option>
                 {corpus?.columns.map((col) => (
                   <option key={col} value={col}>
                     {col}
@@ -172,7 +258,7 @@ export default function Workspace({ project }) {
         )}
       </div>
 
-      {/* Step 2 — construct */}
+      {/* Step 2 - construct */}
       <div className="card">
         <h3>
           <span className="step-badge">2</span>Construct
@@ -184,7 +270,7 @@ export default function Workspace({ project }) {
         <div className="construct-row">
           <div className="grow">
             <select value={constructId} onChange={(e) => setConstructId(e.target.value)}>
-              <option value="">— select construct —</option>
+              <option value="">- select construct -</option>
               {constructs.map((c) => (
                 <option key={c.id} value={c.id}>
                   {c.name} ({c.items.length} items{c.is_seed ? ", library" : ", custom"})
@@ -232,7 +318,7 @@ export default function Workspace({ project }) {
         )}
       </div>
 
-      {/* Step 3 — language, model + run */}
+      {/* Step 3 - language, model + run */}
       <div className="card">
         <h3>
           <span className="step-badge">3</span>Language, model &amp; run
@@ -240,7 +326,7 @@ export default function Workspace({ project }) {
         <p className="hint">
           Embeddings run locally via sentence-transformers; model and language are recorded
           in the run metadata. If the corpus doesn&apos;t match the selected language or the
-          model doesn&apos;t support it, you&apos;ll get a warning — never a silent result.
+          model doesn&apos;t support it, you&apos;ll get a warning - never a silent result.
         </p>
         <div className="run-settings">
           <label className="field language-control">
@@ -380,7 +466,7 @@ function NewConstructForm({ onCreated, onError }) {
         </div>
       </div>
       <label className="field">
-        Scale items — one per line, verbatim from the validated instrument
+        Scale items - one per line, verbatim from the validated instrument
         <textarea rows={5} value={itemsText} onChange={(e) => setItemsText(e.target.value)} />
       </label>
       <button className="primary" type="submit" disabled={saving}>