fix(desktop/transcript): quiescence-based reveal + retired hydration pin, reserved image box (#331)

[agentfleets]

Closes #331

What

Replaces the time-capped hide-until-settled reveal (6b461bb) with a quiescence-based reveal plus a retired hydration pin, and reserves the attachment-image box so blob swaps stop growing rows mid-hydration.

• desktop/src/surfaces/AgentTranscript.tsx
  • Reveal is now driven by quiescence: every totalListHeightChanged re-arms a 150ms quiet timer, so the feed stays visibility:hidden for the whole hydration storm on heavy transcripts instead of popping in mid-storm.
  • The flat 500ms hard cap becomes an 1800ms backstop — it only bounds the hidden window for pathological cases; the quiet timer is the normal reveal path.
  • A settledRef retires the re-pin loop at reveal: post-reveal, pinBottom neither scrolls nor re-arms, so late height changes (font swap, async image pop) no longer cause visible jumps. Bottom-following after reveal remains covered by the existing followOutput + atBottomStateChange contract, unchanged.
• desktop/src/styles/partials/05-transcript-boards.css
  • .md-img-loading now reserves ≈ the final image box (width:100%; aspect-ratio:16/10; max-height:360px) and .md img gains the matching max-height:360px, so the placeholder → blob swap is a geometric near-no-op.

Why

Per the diagnosis posted on the issue (comment thread), the shipped fix fails on heavy transcripts for three reasons: (1) the 500ms cap reveals unconditionally mid-hydration-storm while the 80ms debounce can only reveal early, never delay; (2) the re-pin loop never retires, so every late post-reveal height change is a visible jump; (3) AttachmentImage swaps a 120×80 placeholder for a full-width height:auto image with no reserved dimensions, growing rows by hundreds of px up to seconds later. This PR implements items A+B+C from that diagnosis.

Item D (document.fonts.ready gating) is deliberately skipped: a font swap changes rendered-row heights, which already fires totalListHeightChanged and pushes the quiet timer out — the quiescence window covers it implicitly, so a separate fonts gate would be redundant machinery.

How verified

• npm ci && npm run build (tsc --noEmit + vite build) — clean.
• scripts/lint-desktop-tokens.sh — clean (no new hex / primitive-var / phantom tokens).
• git diff origin/main --stat — only the two intended files; no version bumps or manifest changes.
• CI on this PR is the only device-level verification available in this environment; the WebKit reveal behavior itself wants a manual once-over on a heavy transcript (the scenario in the issue report).

[physercoe]

Reviewed — approve. Focused on what CI can't catch (timer/reveal correctness, CSS geometry).

• Diff matches the description; no hidden changes.
• Timers: reveal() is idempotent (sets settledRef synchronously, then state, then clears both timers); effect cleanup clears both settleTimer and backstopTimer on unmount/re-arm; post-reveal the settledRef guard stops any further re-pin. No leak, no double-reveal.
• No stuck-hidden: the 1800ms backstop is armed unconditionally and calls reveal() regardless of height-change churn, so reveal is guaranteed even if totalListHeightChanged fires forever.
• Bottom-following post-reveal: retiring the re-pin loop relies on followOutput + atBottomStateChange, which covers appended items; an existing bottom row growing after reveal is exactly what the reserved-image-box CSS mitigates — the two halves are complementary by design.
• CSS: .md img { max-width:100%; max-height:360px; height:auto } scales tall images down preserving aspect ratio (no clip/squash, so object-fit correctly unneeded); .md-img-loading is an empty placeholder so its aspect-ratio:16/10 never distorts real content.

Non-blocking: guaranteed-reveal latency for the no-height-change path grew 500ms→1800ms (in practice the 150ms quiet path reveals fast); wide-short real images can still be shorter than the 16:10 placeholder (acknowledged as approximate). Fine.