refactor(combo): decompose combo.ts and enforce the live latency budget

[LMPrado-DZ23]

combo.ts line count

	lines (check-file-size count)
before	4080
after	3779

The frozen entry in config/quality/file-size-baseline.json is lowered 4080 → 3779. It only goes down.

What each new module owns

Module	Owns
open-sse/services/combo/autoCandidates.ts	Building the AutoProviderCandidate rows the auto strategy scores: the connection pool per provider, the per-connection/per-fingerprint target expansion, the blended price per 1M tokens, the latency profile (24h p95 → live metric → per-model bootstrap, plus jitter, error rate and speed telemetry) and the quota picture (reset-window affinity, remaining quota, the opt-in hard cutoff, the soft status penalty). combo.ts re-exports buildAutoCandidates, so every importer resolves as before.
open-sse/services/combo/latencyBudget.ts	Applying the request's maxLatencyMs to live routing: which candidates are routable and which failover targets survive. With no budget both functions are the identity.

exceedsLatencyBudget() was added to open-sse/services/routing/attemptPolicy.ts — the module that already owned the budget machinery — so the decision's latency_over_budget reason and the live filter cannot drift apart.

Equivalence: provider selection is unchanged for requests with no latency budget

Baseline modules extracted read-only with git show origin/release/v3.8.55:… and wired to each other; both halves run against the same shared scoring modules (scoring.ts, engine.ts, taskFitness.ts, decisionStore.ts), the same seeded case generator, a seeded Math.random, a frozen clock, and each in its own process so rotator/self-healing state starts identical.

== PIPELINE GATE base vs head          (real resolveAutoStrategyOrder, 700 seeded cases)
selection equivalence: 700/700 cases identical
Math.random() calls: base=425 head=425 (equal)
per-case Math.random() call counts: all equal

== DECISION-LAYER GATE base vs head    (selectAutoProviderWithDecision + failover ordering)
selection equivalence: 700/700 cases identical
Math.random() calls: base=824 head=824 (equal)
per-case Math.random() call counts: all equal

Sample coverage: 618 selections, 80 strict-budget refusals, 77 rotations, 33 exploration picks, 342 explicit-router picks, 58 early responses, 5155 candidate rows compared field by field (score, eligibility, exclusion reasons, quota, circuit, estimates, factor breakdown). Both halves are self-reproducible (base vs base and head vs head are also 700/700).

The candidate-builder split has its own equivalence run against the pre-split version: 120/120 seeded target-list cases produce byte-identical candidate rows.

== PREVIEW PURITY
previews run: 598
Math.random() calls inside previewRoutingDecision: 0
repeated previews returning the same decision: 299/299
live selections identical with vs without previews interleaved: yes (299 live selections)

Every harness installs a fetch/http trap after all open-sse/ imports (so proxyFetch.ts's import-time globalThis.fetch replacement cannot bypass it), asserts it is the live fetch, and throws on any call. All runs report outbound network attempts: 0. The harnesses exercise selection only — they call resolveAutoStrategyOrder / selectAutoProviderWithDecision / buildAutoCandidates and never reach dispatch; candidates are injected through the buildAutoCandidates dependency.

What live traffic now enforces

• A request that sends X-OmniRoute-Latency-Budget: <ms> gets a RoutingBudget.maxLatencyMs for that request only. No header, no budget — there is no stored combo-level latency cap.
• A candidate whose estimatedLatencyMs (its p95 estimate) exceeds the budget is excluded from selection and dropped from the failover chain, and is reported on the recorded decision with latency_over_budget. Unlike the cost cap, over-budget targets are never merely moved last, so failover cannot walk back over the budget.
• It binds every router strategy (rules and the explicit ones), because the caller asked for it on this request.
• If nothing fits, the request is refused with a 503 naming the budget instead of being served over it.
• Checked per attempt against the candidate's estimate — there is still no cumulative spend check and no elapsed-time check across attempts. docs/routing/ROUTING_CONTRACT.md says exactly this.

Gates

• new-code complexity (the mode CI runs): node scripts/check/check-complexity-ratchets.mjs --base-ref <merge-base> → complexityNewCode=-3 (10 violations in touched files vs 13 on the base), cognitiveComplexityNewCode=-1 (5 vs 6). autoCandidates.ts itself reports 0 violations of complexity, sonarjs/cognitive-complexity and max-lines-per-function. The first push did move the complexity rather than reduce it; the second commit splits the builder into one function per signal, which is what turns the per-file line green.
• four typechecks with tsc directly (open-sse/tsconfig.json, tsconfig.typecheck-core.json, tsconfig.typecheck-api.json, tsconfig.typecheck-dashboard.json) — 0 errors each
• eslint --max-warnings=0 --suppressions-location config/quality/eslint-suppressions.json --pass-on-unpruned-suppressions --no-warn-ignored on every changed file — exit 0
• check-file-size: combo.ts at its new lower number; the only violation is the known pre-existing src/lib/db/core.ts: 1770 > 1745
• mutation-coverage drift (findCoverageDrift called directly): {} over 5080 unit tests, including the new one
• focused routing / auto-combo / combo suites: 189/189 pass. The broad sweep's remaining failures are equally red on release/v3.8.55 (the models-catalog-combo-metadata timeouts swing 5→12 failures run to run on the base tree too, g13-combo-chatcore-golden fails identically on base, and tests/unit/autoCombo/provider-family-combos.test.ts fails 3/3 on base). The repeated SqliteError: UNIQUE constraint failed: call_logs.id in the shard log is pre-existing noise from concurrent call-log writes in combo dispatch tests — nothing in this PR touches call-log persistence, and it is logged, not asserted.

prettier --check still flags open-sse/services/combo.ts — it is not prettier-clean on release/v3.8.55 either (the diegosouzapw#3470 cooldown-retry work deliberately kept ~860 lines at their original indentation), so it is left alone; every other changed file passes.

The diegosouzapw#12600 guard is now behavioural

tests/unit/reset-aware-request-scope-12600.test.ts used to read combo.ts as text and assert getQuotaFetchScope( appeared in it — a proxy that broke the moment the builder moved. It now drives buildAutoCandidates with a spy quota fetcher and caching off: two model families on one Antigravity account produce two quota fetches, two models of one family produce one. The companion "one definition, used by the reset-aware fetcher" assertion stays a source check, because no observable behaviour distinguishes one definition from a duplicate.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: a053fea2-f37c-4a13-9131-756498830b0f

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}