feat(comments): retry failed @claude replies in place

[sam-powers]

What & why

A failed @claude reply used to offer only a single, heavyweight "Re-link session…" button — even when the failure was a transient API hiccup and the right fix was a plain retry. Re-linking opens the session picker and is the wrong tool for a temporary error; there was no way to retry the identical request in place.

This PR classifies the failure and surfaces the affordance that actually fixes it, and stops transient reply state from ever reaching disk.

What changed

• classifyReplyError(message) → { retryable, kind } (co-located with the pure helpers in useClaudeReply.ts, exported + unit-tested). Ordered rules, specific before broad: session → auth → transient → unknown (unknown defaults to retryable). Case-insensitive. The 401 and login auth markers are word-boundary-anchored so a port number like 24010 or a hostname like mylogin.internal can't force a false non-retryable auth verdict.
• retryAIReply(commentId, replyId) (useComments.ts) resets the same reply entry to a fresh pending state (pending: true, error: undefined, text: '') — it never appends a new reply, so the id is stable and the count is unchanged. Unknown replyId is a no-op.
• useClaudeReply.retry(replyId) stashes each ask's inputs in a transient in-memory map, then re-issues the identical spawn reusing the same replyId. A per-replyId generation guard drops late/stale stream events and orphan-cancels a superseded spawn, so a retry that supersedes a slow original isn't clobbered by the original's late done/error. Each spawn cancels by its own captured token, not a replyId re-lookup — otherwise a stale original's late terminal event would cancel the live retry (see the race test). A retryingRef set prevents a double-fire retry. This generation-guard shape is factored to be reused by the upcoming cancel-during-review fix.
• UI by classifier kind (CommentCard.tsx, threaded through CommentLayer.tsx and App.tsx): transient/unknown → Retry primary, no Re-link; session → Re-link primary, Retry secondary; auth → Re-link only (not retryable).
• stripTransientReplyState(comments) drops errored and pending AI replies before serialization, applied at both persistence choke points (the .comments.json sidecar and the draft.json autosave snapshot) so transient UI state never lands on disk. No schema change (SidecarFile stays version: 2).

Testing

• npm run typecheck, npm run lint, npm run format:check — clean.
• npm test — 275 vitest tests pass, including:
  • classifyReplyError — one case per kind, case-insensitivity, ordering (session before transient), and word-boundary negative controls (24010/mylogin don't false-match; a plain word isn't auth; empty/whitespace/undefined → unknown, no throw).
  • retryAIReply — clears error, sets pending, resets text, count unchanged, id stable, unknown replyId no-op.
  • stripTransientReplyState — errored + pending AI replies dropped; user replies and finished AI replies kept.
  • A generation-guard race test (useClaudeReplyRetry.test.ts) that reproduces the cross-generation orphan-cancel hazard: a superseded original's late done is dropped and cancels its own orphan (not the live retry), and a double-fire retry is coalesced to one spawn.
• Rust bar (cargo fmt --check, cargo clippy -- -D warnings, cargo test) unaffected — this is a frontend-only change and stays green.

The full racing-spawn flow end-to-end needs the packaged runtime (a live claude CLI); it's verified manually and the driving window.__quillMock seam is exercised by the vitest race test above. agent-browser is not installed, so no Playwright run here.

🤖 Generated with Claude Code