feat: paint cold transcript tails before full history

[wolfiesch]

What this changes

• Add the negotiated, session-scoped transcript.page read contract with strict entry, cursor, generation, and 512 KiB response bounds.
• Read newest and older transcript pages from bounded JSONL ranges instead of hydrating the whole file. Opaque encrypted cursors stay append-stable and reject replacements, truncation, rewrites, tampering, and cross-session reuse.
• Route paging before the legacy loadSession() path, including transcripts larger than 64 MiB.
• Add a typed client boundary that keeps history cursors separate from the live { epoch, seq } reconnect cursor and rejects stale target generations.
• On a cold session, paint a 64-entry/256 KiB tail before starting the full live attach. Warm cursor attaches remain immediate.
• Add an explicit Load earlier messages timeline action. Older pages prepend by stable entry ID, preserve the visible position, and do not raise the New output pill.
• Document the Flutter cache/anchor handoff and the remaining thin OMP bridge requirement in ADR-014.

Verification

• pnpm check
• host-wire: 60 passed
• protocol: 43 passed
• host-service: 278 passed
• client: 252 passed
• web: 1,147 passed
• The concurrent workspace run reached one existing transcript-search stress timeout at its 5-second limit; the exact test passed in 1.2 seconds on rerun and the complete host-service suite then passed 278/278.
• Commit-time secret scans passed.

Deployment boundary

The source implementation is complete in the T4-owned host packages. The currently released legacy OMP bridge does not consume those packages yet; the planned thin OMP launcher/authority adapter is still required before released clients can negotiate this feature.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.