fix(routing): earliest-reset-first 4-fix bundle (post-deploy hotfix)

[i1hwan]

Summary

• PR feat(routing): earliest-reset-first burn-down strategy + ApexRoute branding #21 (apex-v2.0.0) shipped earliest-reset-first with an additive S = 0.85·T + 0.15·min(Q, 30). Production observation revealed four regressions, all addressed in this single hotfix.
• 48/48 strategy unit tests pass (existing + 18 new RED→GREEN regressions for F1×4, F2×2, F3×6, F4×6). 2791/2791 full unit suite pass. prettier / eslint / typecheck:core / typecheck:noimplicit:core / check:docs-sync clean.
• Plan: .sisyphus/plans/routing-strategy-v6.md (Oracle 4-blocker review + Momus 9-issue review resolved).
• Independent of PR fix(transformer): use streaming TextDecoder so multi-byte UTF-8 survives chunk boundaries #22 (mojibake hotfix) — different files, no logical dependency.

Regressions fixed

#	Symptom (production)	Fix
F1	openai-compatible (self-hosted llama.cpp) → "all candidates excluded" 503	trackScores=0 branches: isAccountQuotaExhausted → excluded; otherwise score=0 fallback (paid still wins, self-hosted usable)
F2	opus/haiku routed to different accounts at the same instant (per-model weekly window split)	Drop modelWindowMapping.ts entirely. Routing reads only overall weekly window. 429s on per-model quota retry via accountFallback
F3	100% fresh Anthropic accounts (Anthropic omits resetAt until first request) auto-excluded	Q === 100 && resetAt === null → max-urgency T=100, score=10000; other null resetAt cases stay missing
F4	APEX 17%/2h outranked GNUMAX 87%/4h (additive formula privileges urgency over abundance)	Multiplicative S = T_pts × Q_remain, score range [0, 10000]. Penalties rescaled 100×. Q_SATURATION_CAP deleted

Hand verification (user's actual scenario)

Account	Track	Q	Reset	T_pts	Score = T·Q
APEX	session	17	2h 23m	40	680
APEX	weekly	35	3d 10h	20	700
APEX	avg				690
GNUMAX	session	87	4h 43m	20	1740
GNUMAX	weekly	59	4d 18h	20	1180
GNUMAX	avg				1460

→ GNUMAX wins by 770 points. Asserted in test F4-1.

Behaviour matrix (post-fix)

Scenario	Result
openai-compatible (cache entry absent)	score=0, eligible; lex tie-break across self-hosted only
Anthropic 429-marked empty cache	excluded: "quota_exhausted_unknown_reset"
paid (score>0) + self-hosted (score=0) mixed	paid wins
paid backoffLevel=4 (finalScore=-8400) + self-hosted	self-hosted wins (intentional)
Q=100, resetAt=null (fresh)	score=10000 (max urgency)
Q=80, resetAt=null (ambiguous)	kind: "missing"
Per-model weekly Omelette=0% (would-have-excluded in v4)	not excluded; routing uses overall weekly only
Q < 5% either track	hard excluded
Ranking tie on T × Q product	earliest-reset asc → lex id asc

API change

• selectByEarliestResetFirst(candidates, sessionId) — modelHint parameter removed.
• scoreAccount(conn), scoreWeeklyTrack(connId), isAffinityValid(conn, sessionId) — modelHint removed.
• auth.ts:706-714 updated; only one caller site, no external API surface change.

Files

• src/sse/services/strategies/earliestResetFirst.ts — rewritten (4-fix bundle)
• src/sse/services/strategies/modelWindowMapping.ts — deleted
• src/sse/services/auth.ts — caller site modelHint removed + comment updated
• tests/unit/auth-strategy-earliest-reset-first.test.mjs — fixture simplified, 18 new tests added
• package.json 3.7.0 → 3.7.1 (patch bump for hotfix)
• docs/openapi.yaml info.version: 3.7.1
• CHANGELOG.md — new [3.7.1] section
• .sisyphus/plans/routing-strategy-v5.md — marked superseded
• .sisyphus/plans/routing-strategy-v6.md — canonical SoT (separate workspace dir)

Out of scope

• Active Sessions fingerprint algorithm (intentional per-agent fingerprinting confirmed).
• cacheTelemetry follow-up fields (still v4 §6 follow-up).
• per-provider strategy override.
• auth.ts.orig cleanup.

Tag plan

Merge → tag apex-v2.0.1 (patch bump on top of apex-v2.0.0).

[Copilot]

Pull request overview

Hotfix update to the earliest-reset-first routing strategy after production regressions in v3.7.0, shifting the strategy to a v6 model that removes per-model weekly window logic and changes scoring to a multiplicative urgency×remaining formulation.

Changes:

• Rewrite earliest-reset-first scoring to T_pts × Q_remain, add fresh-quota handling (Q===100 && resetAt===null), and add a score=0 fallback for no-cache self-hosted providers.
• Remove model-specific weekly-window mapping and drop the modelHint parameter from strategy APIs and the auth.ts call site.
• Add/adjust unit tests for the four production regressions and bump version/docs/changelog to 3.7.1.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/sse/services/strategies/earliestResetFirst.ts	v6 rewrite: multiplicative scoring, fresh quota handling, self-hosted fallback, API signature changes.
src/sse/services/strategies/modelWindowMapping.ts	Deleted; per-model weekly window mapping removed.
src/sse/services/auth.ts	Updates caller to new selectByEarliestResetFirst(candidates, sessionId) signature and updates strategy comment.
tests/unit/auth-strategy-earliest-reset-first.test.mjs	Updates fixtures and adds regression coverage for F1–F4 behaviors.
package.json	Patch version bump to 3.7.1.
docs/openapi.yaml	Updates OpenAPI info.version to 3.7.1.
CHANGELOG.md	Adds 3.7.1 release notes describing the hotfix bundle.

[Copilot]

isAffinityValid can incorrectly keep affinity pinned to a connection that was marked quota-exhausted via markAccountExhaustedFrom429(). In that case both tracks are missing (empty quotas) so this function returns valid: true, while scoreAccount() would exclude it via isAccountQuotaExhausted(). Consider checking isAccountQuotaExhausted(conn.id) here (and returning invalid) so affinity breaks and selection can fall back to another account; add a regression test for this scenario.