fix(sse): exempt tiny-budget reasoning probes and trace persisted-cooldown skips (#12659)

[diegosouzapw]

Closes #12659

Root cause

Two independent gaps in open-sse/services/combo/:

1. Tiny-budget reasoning probes poisoned the model lockout. validateResponseQuality (the combo quality gate) had no way to distinguish a deliberate tiny-budget capability probe (e.g. max_tokens: 1) from a genuine large-budget reasoning exhaustion, so both hit the same reasoning truncated at token limit / reasoning consumed N/M tokens invalid-quality branches — producing a 502 and a recordModelLockoutFailure("quality_failure", 502, ...) call for what was really a capability ping.
2. Persisted-cooldown skips were untraced. executeTargetGates.ts's persisted-connection-cooldown skip branch never called recordComboDecision, persisted_cooldown was not even an allowlisted ComboSkipReason, and buildComboDiag()'s excluded[] only ever sourced from exhaustedProviders/exhaustedConnections — so an ALL_TARGETS_SKIPPED caused purely by persisted cooldowns surfaced as an opaque attempted=0, excluded=[].

Fix

1. validateResponseQuality now exempts a reasoning-truncated response whose actual completion_tokens sits below the fix(api): answer tiny-budget reasoning probes (max_tokens:1) with truncated 200 — no 500 empty response, no cooldown poisoning #10281 tiny-budget threshold (REASONING_BUFFER_MIN_TRIGGER = 256, reused from reasoningTokenBuffer.ts rather than duplicating the constant) — since completion_tokens can never exceed the caller's max_tokens, a tiny count proves a tiny budget was requested without needing to thread the request body through every combo dispatch call site. The exemption returns valid: true, which passes the original upstream 200 straight through (already the correct fix(api): answer tiny-budget reasoning probes (max_tokens:1) with truncated 200 — no 500 empty response, no cooldown poisoning #10281 truncated-200 shape: 200, empty content, finish_reason: length) — no synthetic response needed, and the caller's existing "only lock out on !quality.valid" branch means no model-lockout call happens for a probe.
2. decisionTrace.ts adds persisted_cooldown to the COMBO_SKIP_REASONS allowlist, an optional non-secret detail field on ComboTraceEntry, and a new pure summarizeSkippedTargets() helper that groups skip decisions by reason.
3. executeTargetGates.ts's persisted-cooldown skip branch now calls recordComboDecision(..., reason: "persisted_cooldown").
4. comboAttemptLoop.ts's buildComboDiag surfaces a new skippedTargets[] field (via summarizeSkippedTargets) on the all_targets_skipped diagnostics body. ComboDiagnostics/sanitizeComboDiagnostics in open-sse/utils/error.ts gained the matching optional field, sanitized (length + count capped) exactly like every other diagnostics field — the whole body still goes through buildErrorBody(), never raw err.stack/err.message.

Scope note

The upstream issue's checklist also asks for (a) wiring the same persisted-cooldown trace into roundRobinCombo.ts's dispatch path and (b) threading max_tokens/model context into validateResponseQuality's call sites. I verified by reading source that roundRobinCombo.ts never calls resolvePersistedConnectionCooldownSkipReason at all (round-robin has no persisted-cooldown gate to wire), so there is no equivalent gap there today. For (b), the self-contained completion_tokens-based heuristic reuses the existing regression-test suite's own resolution scale (all #3587 exhaustion cases sit at 512+ tokens, the reported probe at 10) without touching executeTargetAttempt.ts (frozen at its exact LOC baseline) or roundRobinCombo.ts (1 line of file-size headroom) — both zero free bytes I did not want to spend given they're unrelated to the confirmed root cause.

Regression test — RED → GREEN

tests/unit/combo-quality-tiny-budget-probe.test.ts (mirrors the plan's proven repro, the issue's exact reported shape: reasoning consumed 10/10 tokens, finish_reason: length):

• RED (before fix): AssertionError: reproduces #12659: combo validator rejected the tiny-budget probe as a genuine quality failure instead of exempting it like #10281 does on the direct path (reason: reasoning truncated at token limit (finish_reason: length) — no content output)
• GREEN (after fix): valid: true.

tests/unit/combo/combo-skipped-targets-summary.test.ts (new): persisted_cooldown allowlisting, summarizeSkippedTargets grouping/edge-cases, and an explicit assertion that the diagnostics body carries no stack-trace frame and no connection/account-id fragment.

Gates run

• node scripts/check/check-file-size.mjs → OK (138 frozen files, cap 1200 — no growth on touched files)
• node scripts/check/check-complexity.mjs → OK (2798 violations vs baseline 3218)
• node scripts/check/check-cognitive-complexity.mjs → OK (1265 violations vs baseline 1437)
• npm run typecheck:core → exit 0
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> → clean
• npx prettier --check <changed files> → clean
• New tests: node --import tsx/esm --test tests/unit/combo-quality-tiny-budget-probe.test.ts tests/unit/combo/combo-skipped-targets-summary.test.ts → 6/6 pass
• Existing area regression suites (all pass, 129 tests total): tests/unit/combo-quality-validator-reasoning.test.ts (16), tests/unit/combo/combo-decision-trace.test.ts + tests/unit/combo/execute-target-gates.test.ts (15), tests/unit/combo-diagnostics-trace.test.ts (6), tests/unit/combo-stream-readiness-fallback.test.ts + tests/unit/combo-responses-sse-failure-fallback.test.ts + tests/unit/masked-200-exhaustion-fallback-6427.test.ts + tests/unit/combo-terminal-status-policy-10501.test.ts + tests/unit/9303-recovery-hint-all-targets-skipped.test.ts + tests/unit/combo-cooldown-retry.test.ts (55), tests/unit/9303-recovery-hint-all-targets-skipped.test.ts + tests/unit/combo-routing-engine.test.ts (92, overlaps counted once)
• node scripts/check/check-changelog-integrity.mjs → OK

⚠️ base-red inherited: #12732 — unit #12058, integration codex-cache, package-artifact, tarball-smoke, agent-skills-sync

Note

The two PRs the original issue linked as carrying the fix (#12531 and its combo-skip-diagnostics branch, #12661) were both closed without merging — this PR is a fresh implementation pass against the current release/v3.8.51 tip.