bench: regenerate ab-naive / ab-pooled reports (missed in the first pass)

bench: regenerate all 21 courier reports with corrected fan-out + throughput

bench: courier fan-out + work-rate metrics, /trace adoption (test_report_stats.py)

bench: courier fan-out + work-rate metrics, /trace adoption (run.py)

bench: courier fan-out + work-rate metrics, /trace adoption (work_server.py)

bench: courier fan-out + work-rate metrics, /trace adoption (report_html.py)

docs: correct package_courier fan-out + throughput (bug-2273/2274)

bench: courier admission A/B raw data (p10 baseline/P7 + 16-pkg overload raw/P7)

pin: flip live pin to llamafile-v0.10.3+opencoti.c6 + card/USAGE all-compressed matrix + gpu-share

mirror: patches 0138-0144 (backfill c5 gap + c6) + full docs re-sync

release: llamafile-v0.10.3+opencoti.c6 — all-compressed artifact family (5 binaries + 3 DSOs)

pin: flip live pin to llamafile-v0.10.3+opencoti.c6 + card/USAGE all-compressed matrix + gpu-share

mirror: patches 0138-0144 (backfill c5 gap + c6) + full docs re-sync

archive: llamafile-v0.10.3+opencoti.c5 -> releases/archived/ (server-side copy, single commit)

release: llamafile-v0.10.3+opencoti.c6 — all-compressed artifact family (5 binaries + 3 DSOs)

release: llamafile-v0.10.3+opencoti.c5 final — notes + SHA256SUMS + card (win DLL side-load instructions)

pin: add dso win-x86_64 row (c5 DLL)

dso: Windows ggml-cuda DLL (CUDA 13.3, sm_75-120f, compressed fatbin) — 5080 side-load gate PASS

pin: c5 final binary SHAs (bug-2244 sync fix)

c5: gpu-share busy accounting via llama_synchronize (bug-2244) — pa4 + MTP gates pass; patch 0141 recaptured

docs+patches: gpu-share (0141) — USAGE.md, patch series, design doc

pin: c5 re-cut binary SHAs (gpu-share, patch 0141)

c5 re-cut: gpu-share peer coordination (patch 0141) — shm registry, --gpu-share-weight, /gpu/peers, credit/debt duty-cycle pacing

Archive c4 under releases/archived/llamafile-v0.10.3+opencoti.c4/ (per-tag layout, c3 convention)

Archive c4 artifacts to releases/archived/ (repo root holds latest release only)

Upload dso/0.10.3-c5/ggml-cuda-sbsa-aarch64.so with huggingface_hub

Upload dso/0.10.3-c5/ggml-cuda-x86_64.so with huggingface_hub

Upload opencoti-llamafile-0.10.3-c5-x86_64.llamafile.MANIFEST.json with huggingface_hub

Upload opencoti-llamafile-0.10.3-c5-x86_64.llamafile with huggingface_hub

Upload opencoti-llamafile-0.10.3-c5-aarch64.llamafile.MANIFEST.json with huggingface_hub

Upload opencoti-llamafile-0.10.3-c5-aarch64.llamafile with huggingface_hub

Upload opencoti-llamafile-0.10.3-c5-win-x86_64.llamafile.exe.MANIFEST.json with huggingface_hub

Upload opencoti-llamafile-0.10.3-c5-win-x86_64.llamafile.exe with huggingface_hub

Add qwen3-embedding-0.6b-16k embedder llamafile (from claude-hooks v1.4.0, sha 414f6166) for opencoti-server download+install testing (F2 M8 / F4 M8)

package_courier report: 7-column tile grid (14 tiles fit 2 rows)

package_courier: fix sampler (tps_ewma field, run-lifetime sampling, w-mode) + report (x-domain clipping, t0 anchor, aggregate-tps/max-ctx tiles) — bug-2223

bench: A4B Q4 floor-1 control (97/100, StateGraph agent) — floor-1 losses are precision-driven, not harness; correct README claim

readme: A4B Q4 20/25 + F16 ladder rows; F16-vs-Q4 cost + StateGraph zero-loss findings

bench: A4B ladders — Q4_K_M floors 20/25 (cap 64) + F16 full ladder 1/5/15/20/25, all 100/100 zero-loss

bench: defaults — parallel 8, tps-floor 15 scheduler-driven, stall timeout auto-scales with steps (bug-2222)

readme: add 35B-A3B stress-ladder results (100/100 all floors, fan-out knee)

bench: 35B-A3B courier stress ladder (floors 1/5/15/20/25, 100/100 at every rung)

cleanup: drop __pycache__/.ruff_cache/.pytest_cache bytecode + stray integrations/langgraph/bench duplicate

card: LangGraph reference impl is real (StateGraph episodes) + publish integrations/langgraph package

docs: package_courier harness = reference LangGraph implementation

bench/package_courier: true LangGraph StateGraph reference impl (capacity_gate->agent<->tools) + offline routing tests

integrations/langgraph: publish opencoti_langgraph package (reference LangGraph integration, bug-2220)

card: package_courier harness link + release archive layout

bench: publish package_courier harness (report_html.py)

bench: publish package_courier harness (work_server.py)