v0.6.21

• chore(main): release 0.6.21
• test(colpali): cover all three forward paths in #2144 regression tests (review feedback)
• test(colpali,worker): regression tests for #2144 fixes (review feedback)
• fix(colpali): serialize model forwards — transformers output-recorder race (#2144)
• fix(worker): release OOM tracebacks in the no-recovery branch too (CodeRabbit)
• fix(colpali): stop GPU memory accumulation on Vidore evals (#2144)
• test(server): make artifact symlink check deterministic
• feat(server): bind package artifact identity
• fix(evidence): harden worker identity boundaries
• fix(evidence): close worker identity contract gaps
• feat(evidence): propagate worker execution identity
• fix: keep audio prep release pin in sync
• fix(models): restore SPLADE 512-token capacity
• feat(sdk): preserve terminal request metadata on errors
• refactor(server): fold visual-SPLADE serving into the transformers5 bundle
• feat(models): raise qwen36 launch context to 8k
• test(server): resolve OCR model fixtures by file
• fix(models): pin hf_revision for gte-Qwen2-7B remote-code load
• fix(models): serve gte-Qwen2-7B via its own bidirectional forward (#2097)
• test: harden model-filter, feasibility, and dispatch-order coverage
• fix(vision): address review edge cases
• perf(vision): bulk-convert detection outputs
• feat(vision): harden OSS runtime contracts
• fix(extract): close review validation gaps
• fix(extract): filter unconstrained relation endpoints
• feat(extract): add pinned PII quality gate
• test(billing): pin vision reranker token contract
• refactor(billing): scope pair image metering hook
• feat(billing): meter multimodal catalog units
• fix(worker): propagate effective sequence caps
• fix(eval): add faithful Qwen3-VL reference
• fix(reranking): address exact-head review
• fix(server): empty bundle model_filter must advertise zero models, not all
• feat(server): transformers514 bundle + SparseEncoder vision sparse adapter
• fix(keda): keep gateway coordination under autoscaling
• fix(keda): verify lean forward upgrade
• fix(keda): simplify forward OTLP upgrade
• fix(model): restore SPLADE quality sequence cap
• fix(cloud): clarify idempotent deletion tests
• feat(cloud): enforce exact managed catalog authority
• fix(keda): support forward-only OTLP upgrade
• fix(bench): capture generation timeout evidence
• fix(server): remove unused fake adapter import
• fix(metering): reject invalid score counts
• fix(cloud): address billing review findings
• fix(reranking): emit authoritative fake usage
• refactor(reranking): keep fake adapter outside scope
• style(core): format compatibility assertions
• fix(core): preserve unit count wire order
• feat(sdk): expose terminal billing metadata
• feat(core): report authoritative score pair usage
• feat(core): expose authoritative request usage
• fix(reranking): address exact-head review
• fix(reranking): reject blank compatibility inputs
• fix(reranking): emit fake score usage
• feat(gateway): add strict Cohere rerank compatibility
• feat(reranking): harden Qwen score runtime
• feat(candle): add native SPLADE sparse encoding
• fix(audio): sandbox image builds the audio wheel opportunistically
• feat(models): onboard Snowflake/snowflake-arctic-embed-s (scaffolded)
• fix(cloud): pin governed catalog weights
• fix(inference): harden abandonment cancellation
• fix(inference): cancel abandoned non-generation work
• fix(audio): address exact-head review findings
• fix(audio): harden release integration
• test(audio): isolate optional compatibility deps
• fix(audio): harden Whisper compatibility
• feat(audio): add native Whisper transcription
• fix(transport): address chunk protocol review
• test(telemetry): assert canonical worker profile identity
• refactor(telemetry): expose one gateway metrics facade
• fix(telemetry): stage safe KEDA rollback migration
• fix(models): bound Qwen grammar admission
• fix(models): address Qwen profile review
• fix(models): promote Qwen3.5 A100 default
• test(server): tighten queued option assertions
• feat(transport): harden local binary ingress
• fix(telemetry): close rollback and audit gaps
• fix(ocr): distinguish missing document inputs
• feat(ocr): add document-aware extraction contracts
• fix(generate): preserve governed streaming overrides
• feat(generate): govern runtime options and streaming
• fix(telemetry): preserve delta metrics across backends
• fix(dashboards): cover result chunk telemetry
• refactor(transport): harden generic chunk delivery
• fix(serving): reject unsigned generation seeds
• fix(api): align seed validation across local clients
• fix(serving): correct generation seed contract
• fix(cloud): avoid serving sync for pin checks
• fix(inference): honor native profile selection
• fix(serving): normalize direct-route seeds
• fix(serving): preserve seeded SGLang streams
• feat(inference): add bounded 8k ColBERT scoring
• fix(serving): isolate SGLang vision extraction workers
• feat(retrieval): complete sparse and multivector launch support
• feat(extract): add GLiNER2 structured and relation support
• fix(gateway): preserve encode output dtype
• fix(sidecar): preserve raw encode ids and cap results
• fix(candle): make multivector profile contract explicit
• refactor(telemetry): enforce single-emission routing
• test(gateway): configure fake stack KEDA lane
• data(bench): re-floor Splade_PP_en_v2 at the vanilla numbers (#1919)
• feat(telemetry): centralize observability on OTLP
• perf(candle): enable Arctic XLM-R fp16 fast path
• perf(sidecar): coalesce saturated scheduler tails
• feat(sidecar): expose scheduler request latency
• perf(server): reduce BGE packing synchronization
• perf(server): batch BGE-M3 tokenization
• fix(sidecar): avoid double-counting adaptive latency
• feat(telemetry): cover managed service signals
• feat(telemetry): codify dashboards and edge signals
• feat(telemetry): add safe request logs and latency
• feat(telemetry): add Better Stack OTLP pipeline
• fix(cloud): close i6pn lifecycle review gaps