feat(tui): brand the startup transcript with a quiet session header

[TheGreatAxios]

Summary

• The startup transcript opens with the Corbits Code wordmark and a single quiet essentials line (profile, model, effort, permission mode)
• The landing lockup still plays during world load, then clears to the branded header ahead of the other deferred startup notices
• Startup echoes (picker choice, permission notices) still land in the transcript but collapse when identical back-to-back, so repeats read once; the prompt border label gains a yolo mirror of the permission mode
• The header is a startup snapshot: later /yolo or effort toggles move the prompt border label only

Verification

• bun run check (typecheck, build, test) green
• New tests cover the session header composer, the yolo mode segment, and the dedupe/flush-first transcript behavior

Fixes CL-7938

[linear-code]

CL-7938

[TheGreatAxios]

Critic review · Request changes

Startup branding reads right at the behavior level: the header surfaceSystemNotice call sits immediately before the fire-and-forget block in wirePostStartup, so FIFO deferred order puts the header first; the mode segment is plumbed at all three live label sites (initial modelLabel, effort toggle, /yolo toggle — the effort toggle previously would have dropped mode); and the observe sequence guard in chrome.ts is sound (child rows share the module-level paint sequence via paintAppendStreamRow, so intervening child paint breaks the seq - 1 adjacency). Two issues before merge:

Blocking

• src/tui/components/session-header.test.ts ("wordmark alone when no essentials apply", "wordmark leads the quiet essentials line") pins the literal wordmark "corbits code" instead of importing LOCKUP_WORDMARK (src/tui/lockup.ts:42). Per the AGENTS.md test rule ("assert the contract instead… This bar is a review and authorship rule"), a brand rename breaks these tests with zero behavior change. Assert composition against the imported constant.
• src/tui/startup-transcript.test.ts pins other modules' copy: "Chose muse-spark." (model-picker wording) across the three dedupe tests, and "Permission prompts are disabled by your saved default (/yolo off to re-enable)." (owned by src/tui/runner/wiring.ts) plus the literal header string in the flush-order test. The contracts under test are duplicate-collapse and FIFO flush order — both hold with synthetic strings (build the header row via composeSessionHeader if the integration value is wanted).

Should-fix

• src/tui/shell/chrome.ts (isDuplicateSystemEcho, new): equality is role + text only, ignoring agent (writer) and meta. Non-user rows feed writer-voice tracking (src/tui/shell/transcript.ts:58-63) — verified against current code that two same-text system rows from different writers flip the transcript to multi-voice. Back-to-back same-text system rows from different writers (multi-agent observe traffic) would collapse the second row and skip noteAgentVoice, losing the second-voice relabel. Cheap fix: include agent (and meta) in the comparison, plus a regression test with same-text / different-writer rows.

Nits (non-blocking)

• prompt-action-bar-label.ts: mode?: string | undefined — the | undefined is redundant with ?.
• session-header.ts: const WORDMARK = LOCKUP_WORDMARK adds a hop; fine either way.
• The dedupe applies to all runtime system rows, not just startup echoes — accepted trade-off, noted so future repeating notices aren't a surprise.

Verdict

Request changes: fix the copy-pinning tests and the writer-blind dedupe comparison. No other findings in the wiring/chrome paths reviewed.