fix(harness): bounded retry + honest turn end for empty final steps

[itsdestin]

What

Native turns were silently ending mid-task when the model's final step came back with zero content (no text, no tool calls) under a valid stop finish — experienced as ~3 minutes of nothing after a tool result, then an unexplained stop requiring a manual "continue" (observed 3x live on stealth/ox-alpha, 2026-08-20/21; also reproduced on qwen3.8-max).

Now:

• An empty step with an orderly finish reason gets one silent re-run (history untouched — the empty step never pushed anything, so the request is repeatable; both attempts bill into turn usage).
• A second consecutive empty step ends the turn honestly with a new stopReason: 'empty_response' and a visible footer: "The model returned an empty response. Retrying may help."
• finishReason-gated: 'length'/'content-filter' empties keep today's exact max_tokens/refusal behavior — the retry never masks a real stop reason.
• Fully-silent turns render the footer too (spec decision 4): the reducer creates the turn on abnormal-stopReason turn-complete, and the bubble renders a footer-only row — previously this worst-case shape had nothing to attach the explanation to.

What it deliberately does NOT touch

Stall clocks, the park guard, mapStopReason, the frozen emit surface (no new event types or IPC — the state rides the existing turn-complete payload), TRANSCRIPT_INTERRUPT, Kotlin (native stopReason is opaque passthrough — verified).

Docs

• Spec: youcoded-dev/docs/active/specs/2026-08-21-empty-final-step-turn-recovery-design.md (reviewed + ratified; includes prior-art comparison with opencode v1.18.21's sibling unknown-finish fix)
• Plan: youcoded-dev/docs/active/plans/2026-08-21-empty-final-step-turn-recovery-plan.md (independently reviewed; verdict approve-with-fixes, findings folded in)

Testing

TDD throughout — 15 new tests written failing-first: the retry ladder (cases 1–4), reasoning-only steps classified empty (case 5, new reasoningChunks helper), the finishReason gate (case 6), interrupt-wins (case 7), specialist-child settle (8a/8b), footer copy (case 9), suite-stability (case 10), and the fully-silent-turn reducer/bubble pair (case 11). Downstream eval-runner/specialist suites that staged "model simply stopped" with one empty step doubled it to keep their premise under the new contract. harness-stall-watchdog.test.ts untouched and green; verify.sh green (tsc, vitest, knip, eslint, ast-grep); full desktop suite 5,995 passed.

Known open item: the footer-only row (decision 4) has not been visually inspected — it renders only in the rare all-empty failure state; styling can be tweaked post-merge if needed.

🤖 Generated with Claude Code

[itsdestin]

Deep review + fixes (2026-08-21, third session)

Two independent adversarial review rounds ran against this branch; every finding was verified against the code before fixing, and each fix round was itself independently re-reviewed before pushing. Five new commits (3270a00a..fe514664):

The big one

The PR's user-visible fix was dead code. Both timeline render sites (ChatView.tsx, buddy BubbleFeed.tsx) drop segment-less turns before AssistantTurnBubble ever mounts — and every shipped test mounted the bubble directly or asserted on reducer state, so all tests passed while the real app still showed the exact unexplained silence this PR exists to fix. Fixed with one shared predicate pair in chat-types.ts (abnormalStopReason / shouldRenderAssistantTurn) used by the reducer's mint gate, the bubble's footer gates, and both timeline gates — pinned by a new test that mounts ChatView itself (state in → footer out).

Also fixed

• Whitespace-only steps were pushed to history (truthiness gate) yet classified empty (trim gate) — the retry then re-sent a conversation ending in a dangling whitespace assistant message (400 on Anthropic-shaped endpoints). One shared stepHasText predicate now feeds both gates; the interrupt path trims too.
• Ghost turns on replay: the segment-less mint broke turn-complete's absorb contract — every re-dock replay appended a fresh ghost turn. Abnormal completions now record their uuid in seenUuids; the mint checks it. Minted turns take the event's timestamp; the two metadata-stamp sites merged into one (they had already diverged on model).
• Empty 'tool-calls' finish (announced call dropped as malformed — the likeliest shape on small local models) now gets the retry instead of surfacing the raw passthrough reason; the orderly-finish list is now ORDERLY_EMPTY_FINISHES, one const next to mapStopReason.
• Orphaned 'Preparing…' card during the retry: withdrawn explicitly (same rule as the manual/stall retry paths), via a new StepResult.pendingPreparing field. Pinned by test.
• Diagnosability: console.error → structured log() (~/.claude/desktop.log).
• Copy: 'The model returned an empty response twice. Retrying may help.' — states the verified fact; the footer-only row also gained the timestamp every other message has.
• docs/chat-reducer.md no longer contradicts the code (stopReason open set; segment-less turns exist).

Visible side effects to be aware of (deliberate, flagged)

• Interrupting while a tool call is still 'preparing' now leaves a small 'Interrupted.' footer where the turn previously vanished entirely.
• A thinking-only CC turn ending max_tokens/pause_turn now renders its footer where there was silence. pause_turn is the one to watch (a standing row while CC continues in a new turn) — easy to narrow later if it annoys.

Known residuals (documented in workspace ROADMAP, deliberately not fixed here)

• Whitespace-only steps' persisted deltas still rebuild into resumed history (cosmetic; smaller than pre-fix; proper fix touches the history-rebuild parity arbiter).
• The stamp path (not the mint) is still not uuid-idempotent when a replay lands mid-live-turn — pre-existing.
• A non-empty step with one completed + one dropped tool call still strands a preparing card until turn end — pre-existing, now has a cheap fix shape via pendingPreparing.

Verification: bash scripts/verify.sh green (tsc, related vitest, knip, eslint, ast-grep); 156 tests across the four touched suites.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lh1iH7j6qQDcHRFWMS1dhL