QVAC-17940 chore[skiplog]: backmerge release-sdk-0.11.0 — version bump, changelog, NOTICE, tooling

[NamelsKing]

Companion release PR

• QVAC-17940 chore[skiplog]: release sdk 0.11.0 #2090

This PR lands the same release metadata on main per gitflow.md "Keep main aligned". Content is identical to PR #2090 (same single cherry-picked commit). Tagged [skiplog] so it does not appear in future changelogs.

---

🎯 What problem does this PR solve?

@qvac/sdk is ready to cut v0.11.0. 29 PRs have landed on main between the sdk-v0.10.0 tag and the current release-sdk-0.11.0 tip (e282e704), and the package version bump + changelog set need to land on the release branch so the publish workflow can ship to npm.

The release completes the request-lifecycle and cancellation overhaul that began in 0.10.0: every long-running SDK call now flows through a unified RequestRegistry, exposes its requestId synchronously on the returned promise, and can be cancelled individually with cancel({ requestId }). The cancel(...) wire envelope is consolidated to two shapes, two legacy call signatures are removed (cancel({operation:"downloadAsset", downloadKey}) and cancel({operation:"rag", workspace})), and the SDK gains typed instanceof for policy/cancel errors across the RPC boundary.

Alongside the lifecycle work the release adds: Harmony / Qwen3.5 / Gemma4 tool-call dialects, FLUX.2 multi-reference fusion + per-call LoRA on diffusion, ESRGAN upscaling (post-step + standalone upscale() API), Whisper VAD + end-of-turn events, multi-GPU split-mode / tensor-split / main-gpu on LLM and embed, a reasoning_budget knob for thinking models, and a Parakeet 0.4.0 GGUF backend with duplex streaming.

Full per-PR breakdown in changelog/0.11.0/CHANGELOG.md; human-readable narrative in CHANGELOG_LLM.md; breaking-change migration guides in breaking.md; API code samples in api.md.

📝 How does it solve it?

Release metadata:

• Bump packages/sdk/package.json from 0.10.2 to 0.11.0.
• Add packages/sdk/changelog/0.11.0/ with:
  • CHANGELOG.md — sectioned entries (✨ Features, 🔌 API, 🐞 Fixes, 📦 Models, 🧪 Tests, 🧹 Chores, ⚙️ Infrastructure), one bullet per valid PR (29 total after filtering [skiplog] / invalid-format / revert / exclusion-listed).
  • breaking.md — auto-extracted BEFORE/AFTER for the three [bc] PRs (QVAC-17869 feat[bc]: migrate SDK parakeet plugin to 0.4.0 GGML + duplex streaming #2018 parakeet GGUF migration, QVAC-18827 feat[bc|api]: add unloadModel autoClose option, default-off on Bare #2024 unloadModel autoClose, QVAC-18183 feat[bc]: CLI cancel bridge + cancelHandler retirement #2074 cancel-shape removals).
  • api.md — auto-extracted code samples for the 13 [api] PRs.
  • models.md — added/removed model constants (Parakeet GGUF additions, Marian Opus NMT removals).
  • CHANGELOG_LLM.md — hand-authored narrative release notes following the v0.10.0 shape.
• Rebuild the root packages/sdk/CHANGELOG.md aggregate (15 versions, 0.11.0 first; prefers per-version CHANGELOG_LLM.md).
• Regenerate packages/sdk/NOTICE against the current packages/sdk production dependency tree (178 JS deps + 209 models).

Tooling improvements bundled with the release (surfaced during changelog generation, kept together for a single release-line touch — see "🧪 How was it tested?" for the specific failure modes that motivated each):

• scripts/sdk/generate-changelog-sdk-pod.cjs:
  • Crash guard — processSDKPRs no longer throws Cannot read properties of undefined (reading 'tags') when a PR title matches a validator exception (Merge / Revert / Squash / version-bump / release(...)). Previously fatal; now warned and skipped. Triggered in this release by PR Revert qvac 17869 parakeet ggml #2059 (Revert qvac 17869 parakeet ggml).
  • Per-package PR exclusion table (EXCLUDED_PRS) for already-merged cross-cutting chores that incidentally touch a package path. Seeded with sdk.1860 (monorepo-wide path simplification across 26+ packages — only incidentally touched packages/sdk via a test executor and a historical changelog doc).
  • Short announcement-post.txt format — generateAnnouncementPost rewritten to emit a Slack-ready post containing header + 3 links + optional breaking-changes block + thanks footer, matching the team's new convention (mirrors the 0.10.0 Slack post). Dropped ~180 lines of dead helpers (parseChangelogMarkdown, formatAnnouncementBullet, CHANGELOG_HEADING_TO_KEY, SLACK_SECTION_HEADINGS, MAX_ANNOUNCEMENT_BULLETS). Breaking-changes detection now keys on breaking.md file presence (structurally direct — if there's a migration guide, link to it) instead of re-parsing CHANGELOG.md.
• .cursor/skills/sdk-changelog/SKILL.md — Step 5 updated to document the new short format and the file-presence detection contract.

🧪 How was it tested?

• node scripts/sdk/generate-changelog-sdk-pod.cjs --package=sdk — exited cleanly after the bypass-validation crash guard + QVAC-16441 feat: simplify package folders, files and paths in the monorepo #1860 exclusion landed; fetched all 47 candidate PRs, filtered to 29 valid, generated CHANGELOG.md / breaking.md / api.md / models.md and rebuilt the root aggregate.
• node scripts/sdk/generate-changelog-sdk-pod.cjs --package=sdk --update-root-changelog — picked up the hand-authored CHANGELOG_LLM.md and re-rendered the root aggregate (root prefers LLM file when present).
• node scripts/sdk/generate-changelog-sdk-pod.cjs --package=sdk --generate-announcement-post — emits the new short format. Cross-validated against the v0.10.0 Slack post screenshot: header / 3 links / breaking-changes block (gated on breaking.md presence) / footer. Smoke-tested the no-breaking branch by re-generating against v0.9.1 (which has no breaking.md) — confirmed the breaking-changes block is correctly skipped.
• node .cursor/skills/notice-generate/scripts/generate-notice.js sdk — license-checker against packages/sdk/node_modules (178 production JS deps) + model attributions from models.prod.json (209 models). Sorted deterministically; re-runs produce byte-identical output.
• Two PRs hit format-validator warnings and were surfaced but not blocked (per skill: "warned, not silently dropped"): chore[mod]: update registry model constants #1948 ([mod] tag requires a Models section in the body — missing) and chore[notask] sdk: upgrade ocr-onnx and translation-nmtcpp #1999 (extra space before colon in title). Both excluded from the changelog with explicit warnings; will be flagged to PR authors for fix-and-backport if desired.
• No source-code changes in packages/sdk/ itself — release content + tooling improvements only.

💥 Breaking Changes

Three [bc] PRs are folded into this release (full BEFORE/AFTER in changelog/0.11.0/breaking.md):

1. unloadModel no longer auto-closes the Bare worker (QVAC-18827 feat[bc|api]: add unloadModel autoClose option, default-off on Bare #2024). Default autoClose flips by runtime: true on Node/Electron (unchanged behaviour), false on Bare. Pass autoClose: true | false to override explicitly.
2. Parakeet plugin moves to the 0.4.0 single-file GGUF API (QVAC-17869 feat[bc]: migrate SDK parakeet plugin to 0.4.0 GGML + duplex streaming #2018). Multi-file parakeet*Src config fields removed; the addon auto-detects TDT / CTC / EOU / Sortformer from GGUF metadata. Legacy multi-file constants (PARAKEET_TDT_ENCODER_INT8, PARAKEET_TDT_DECODER_INT8, etc.) replaced by single GGUF constants (PARAKEET_TDT_0_6B_V3_Q8_0, PARAKEET_CTC_0_6B_Q8_0, PARAKEET_SORTFORMER_4SPK_V1_Q8_0, PARAKEET_EOU_120M_V1_Q8_0).
3. Two legacy cancel(...) call shapes removed (QVAC-18183 feat[bc]: CLI cancel bridge + cancelHandler retirement #2074). cancel({operation:"downloadAsset", downloadKey, clearCache}) and cancel({operation:"rag", workspace}) removed because neither carried a requestId. Migrate to cancel({requestId: op.requestId}) via the decorated promise (primary path) or cancel({modelId, kind: "rag"}) broad-cancel (escape hatch). Every other cancel(...) shape is preserved by client-side normalization.

🔌 API Changes

13 [api] PRs land on the public SDK surface — full per-PR code samples in changelog/0.11.0/api.md. Highlights:

• Request lifecycle + per-request cancel (QVAC-18181 feat[api]: introduce request lifecycle primitives with signal-based cancel #1949, QVAC-18182 feat[api]: typed cancel outcomes on the wire + atomic KV-cache via KvCacheSession #2007, QVAC-18183 feat[api]: cancel capability + per-handler cancel scope + structured logging #2036, QVAC-18183 feat[api]: inference-handler migrations #2058, QVAC-18183 feat[api]: non-inference migrations + decorated-promise requestId #2060, QVAC-18183 feat[bc]: CLI cancel bridge + cancelHandler retirement #2074) — every long-running call (completion, loadModel, embed, transcribe, transcribeStream, translate, finetune, downloadAsset, and the three cancellable RAG ops) returns a decorated promise with synchronously-readable requestId. New RequestRejectedByPolicyError (52420) and InferenceCancelledError (52419) round-trip across the RPC boundary with their typed fields intact.
• Multi-GPU support on LLM + embed (QVAC-17487 feat[api]: expose split-mode and tensor-split in SDK LLM config #1759) — split-mode, tensor-split, main-gpu on LLM (hyphenated); splitMode, tensorSplit, mainGpu on embed (camelCase).
• Diffusion enhancements (QVAC-18135 feat[api]: add FLUX.2 multi-reference fusion and LoRA adapter support to diffusion API #1838, QVAC-18475 feat[api]: add ESRGAN upscale support to SDK diffusion #1930, QVAC-18741 feat[api]: add standalone image upscaling support to the SDK #1990) — FLUX.2 multi-reference fusion (init_images: Uint8Array[]), per-call LoRA, ESRGAN as post-step or standalone upscale() API.
• Whisper VAD + end-of-turn events (QVAC-17282 feat[api]: expose whisper VAD and end-of-turn events in transcribeStream #1848) — transcribeStream({ emitVadEvents: true }) opt-in mode yields a discriminated event stream with live VAD probabilities + turn boundaries.
• New tool-call dialects (QVAC-17324 feat[api]: add harmony tool-call dialect (gpt-oss) #1878, QVAC-18717 feat[api]: add Qwen3.5, Gemma4 tool-call dialects and reasoning_budget param #1974) — Harmony (GPT-OSS), Qwen3.5 Pythonic-XML, Gemma4 native; reasoning_budget knob for thinking models.
• Standalone CLI bundle verification (QVAC-18679 feat[api]: wire qvac verify bundle into Expo plugin #2000) — Expo prebuild now runs qvac verify bundle against the worker bundle before the SDK dist/ copy; new BundleVerificationFailedError (50609).

[NamelsKing]

/review

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*

[NamelsKing]

/review