feat: refuse low-information single-token extractive spans (confidently-wrong QA pattern)

fix: raise fast-tier budget gate to 140ms so extractive answers always finish inside the 200ms cap

feat: strict 200ms latency mode - deadline on by default, pre-fast-tier budget gate, post-STT deadline scope for voice

fix: nest chain budget inside outer generation deadline to bound recovery double-pass

fix: script-aware grounding - cross-lingual answers grounded on number conservation; Indic-aware tokenization

fix: retry truncated Sarvam JSON within attempt budget and raise its token ceiling

fix: tolerate proxy-injected request ids; document generation chain and measured latency

feat: sarvam-first generation chain (groq throttles under sustained load)

fix: groq connection pool starvation - multiple pool slots and bounded attempts

feat: chain-wide generation budget so total latency stays bounded across tiers

feat: borderline evidence goes to grounded generative review; bounded API client timeouts

feat: route English queries straight to generative chain (corpus is Indic-only)

chore: drop pycache artifacts from repo

fix: self-healing Groq payload on HTTP 400 and detailed chain attempt diagnostics

feat: ordered generation chain (groq -> sarvam -> local GGUF) with per-tier deadlines and bounded llama.cpp threads

fix: compile GBNF answer grammar to LlamaGrammar for llama-cpp-python 0.3.35

fix: resilient generation - ZeroGPU circuit breaker, resident CPU model, API fallback net

feat: strict-latency deadline guard, grammar-constrained GGUF decoding, hardened guardrails (#28)

fix: hard 30s generation timeout and slow-first-attempt recovery guard (#27)

fix: drop obsolete ambient reset rule for fixed background (#26)

fix: generate multilingual answers instead of extracts (#23)

fix: export Qwen model for Space startup (#21)

Fix multilingual retrieval and Qwen fallback (#20)

Rebase retrieval latency optimization onto latest Space (#18)

Fix selected output language end to end (#17)

Make the UI visibly clearer and more useful (#13)

Improve answer UX and provenance transparency (#11)

Honor selected answer language (#9)

Remove slow GGUF JSON grammar and shrink multilingual prompt (#8)

Use Blackwell cu130 Q4 runtime and short multilingual generation path (#7)

Fallback to CPU llama.cpp when ZeroGPU CUDA GGUF kernels fail (#6)

Run Qwen3.5 0.8B Q4_0 locally on ZeroGPU (#5)

Pin circles with fixed viewport backgrounds (#4)

Lock background circles to viewport root (#3)

Fix fixed background circles and stable input toggle (#2)

Fix multilingual evidence corroboration (#1)

Restore dropdown and ambient styling

Restore text mode and language menu

Fix blank screen during model startup

Allow hosted UI embedding

Add original Anvaya UI and workflow