Prepare for tinycloud 0.3.12: watch speech segments fix + --transcript/--content (skill floor → 0.3.12)

[kdr]

Prepares the distribution repo for tinycloud 0.3.12 (binary built from the private source repo; release build 0.3.12+fd38d41).

Important

Do not merge until the CDN serves 0.3.12 (channels.stable = 0.3.12, generate-manifest.mjs --check --version 0.3.12 green). This release raises the skill floor to 0.3.12, so merging early would make preflight hard-fail every existing install.

What 0.3.12 ships (binary)

• Critical watch fix: watch --speech-only / audio sources returned segments: []. Segments (and embedded shots) now inline their overlapping utterances as speech: string[]; sources with speech but no visual segmentation get uniform 20s segments synthesized ("segmentation": "uniform:20"); pre-0.3.12 caches heal automatically (persisted on watch cache hits, in-memory during search scans). Un-breaks search --field speech over cached watches. New feature id watch.speech.v1.
• watch --transcript and watch --content are now wired (previously declared in commands --json but no-ops): data.transcript (speaker-labeled when multi-speaker) and data.content (full describe markdown), derived per invocation, never persisted, served on cache hits.
• Breaking: watch --json-index removed — now errors as an unknown flag (its consumer was deleted pre-0.3.0; it did nothing).
• Background watch jobs reach foreground parity (sidecars, transcript/content on reconcile).

Contract delta vs 0.3.11 (verified against the real build): verbs stay 16, features 32→33 (+watch.speech.v1), commands --json diff touches only watch (-json-index), tarball members still 84.

Dist repo changes

• Version parity → 0.3.12 (5 spots): package.json, .claude-plugin/marketplace.json ×2, .claude-plugin/plugin.json, tinycloud-skill.json skill_version; README npx/install pins; CLAUDE.md examples + release narrative + CDN prose.
• Skill floor raised (the 0.3.7 floor-raise pattern, since the speech fix is load-bearing): min_version 0.3.7 → 0.3.12, supported_range >=0.3.12 <0.4.0, watch.speech.v1 added to required_features — mirrored in preflight.sh (MIN_VERSION, REQUIRED_FEATURES; 24 ids, CI parity job green).
• Docs: verbs.md watch section — dropped [--json-index] from the signature, documented speech-inlined segments / uniform:20 synthesis / cache heal and the now-working --transcript/--content; search section notes --field speech over cached watches; glossary.md watch-context entry mentions inlined speech[].
• Unchanged on purpose: ci.yml (smoke matrix is dynamic; the frozen 0.3.0 leg's preflight assertion is version-aware and now expects exit 11), tests/fixtures (frozen 0.3.0 placeholders), install.sh, lib/*, workflow skills.

Verification (against the real 0.3.12 build)

• npm test: 42/42 offline and 42/42 with TINYCLOUD_TEST_TARBALL pointing at the real tarball.
• scripts/smoke-test.sh with EXPECTED_VERSION=0.3.12: all pass.
• Preflight floor proof: 0.3.12 binary → exit 0; 0.3.11 binary → exit 11 ("below required 0.3.12").
• preflight ↔ tinycloud-skill.json parity: 24 feature ids, MIN_VERSION match (mirrors the CI check).
• shellcheck (install.sh, smoke-test.sh, preflight.sh) and claude plugin validate . clean.
• Doc claims grounded in commands --json: watch has transcript/content, no json-index.

---

Note

Medium Risk
Raising the skill floor to 0.3.12 will reject all users still on 0.3.11 once merged; timing must match CDN availability. Docs also advertise a breaking CLI change (watch --json-index removed in the binary).

Overview
Prepares the public distribution repo for tinycloud 0.3.12: npm package, Claude plugin metadata, and skill manifest all bump to 0.3.12, with README/CLAUDE.md examples and CDN/smoke pins updated to match.

Because the 0.3.12 binary fixes load-bearing watch speech behavior (speech[] on segments, uniform 20s synthesis for speech-only/audio, cache heal, search --field speech, feature watch.speech.v1), the agent skill minimum version is raised from 0.3.7 → 0.3.12 (tinycloud-skill.json supported_range, preflight.sh MIN_VERSION, and watch.speech.v1 added to required_features). Do not merge until CDN channels.stable serves 0.3.12 — otherwise preflight will hard-fail existing installs below the new floor.

Skill reference docs are aligned with the new contract: verbs.md drops watch --json-index, documents inlined segment speech, --transcript/--content, and speech search; glossary.md notes speech[] on watch context. No launcher/install/CI workflow code changes beyond version strings in maintainer docs.

^{Reviewed by Cursor Bugbot for commit fdbfa52. Bugbot is set up for automated code reviews on this repo. Configure here.}