fix(opencode): keep the target format and client session per request, not on the shared executor

[maxmad64bis]

Stacked on #14148. Only the last commit is this PR's; I'll rebase once it lands.

⚠️ base-red inherited: #13866

Summary

Default change: none. Behavior change: a JSON caller of a Responses model no longer gets the raw event stream back when another request finishes while it waits. The shared executor kept target format and client session as instance fields written at request start and read after later awaits, so a slow Responses JSON request got the event stream (event: response.completed …) once a Chat request finished first; both values now live in a per-execute() context, with the public fields unchanged outside execute().

Related Issues

• Related to fix(opencode): replay a request in the other tool shape when its tools are refused #14148

Validation

• Change type: provider
• Focused tests and category gates from the golden path
• npm run lint — ESLint on the touched files is clean; the full run is red on the base (5 errors in files this PR doesn't touch)
• Reconciled with the current active release base; focused checks rerun afterward
• Production-code changes include a new or updated automated test in this PR

Tests Added Or Updated

• tests/unit/opencode-request-format-race.test.ts (new, 4 cases): slow Responses JSON caller with a Chat request finishing meanwhile (fails before the change), two in flight each keeping session and format, a request ending not clearing what another still uses, plain fields unchanged outside execute().

Coverage Notes

• opencodeRequestContext.ts (new) and the accessors in opencode.ts: mutating the code (format or session always in the shared field, execute without a context) makes the named tests fail. The 41 test files that touch the executor pass unchanged. No coverage moved down.

Reviewer Notes

• It depends on fix(opencode): replay a request in the other tool shape when its tools are refused #14148: the wrapper it hooks into, withRequestShapeRetry around executeOnce, is introduced there. The public _requestFormat and _clientSession stay as accessors: inside execute() they use the request's own context, outside it plain fields, keeping the ten existing tests that assign them unchanged.
• The Muse Spark output-budget helpers move to opencodeMuseSpark.ts unchanged and re-exported, taking opencode.ts from 1195 to 1125 lines; isResponsesTerminalLine is now exported and reportedCompletionTokens leaves normalizeMuseSparkFinishReason, putting its complexity under the limit. Not changed here: the account list and cursor are also rebuilt per request on the shared instance (same class of problem), and reads of _requestFormat after a request ended still see the cleared value, as before.
• CI reds are inherited from the red base (🔴 Release branch not green: release/v3.8.51 #13866), not from this diff: the third-party PR fix(sse): treat antigravity empty completions with a normal stop as valid 200s (#14160) #14243 on the same base release/v3.8.51 fails the same 9 jobs (API Route Typecheck, Docs Gates, Fast Quality Gates, Merge integrity, ESLint, Unit fast-path 1-4/4); every file cited by the failing gates is outside this diff. Non-blocking for this PR.

[diegosouzapw]

Thanks @maxmad64bis — merging via the release merge-train. Validated in local merge-train (mt-train10c) on the devbox @ train tip 4d841aa1c740bbaa03868dc0a403c62099a99a42 with the 72 sibling PRs of this batch: typecheck:core, file-size, complexity, cognitive-complexity, changelog-integrity green; changed-area node:test 831/831 (0 failing) and vitest 480/482 — the two reds are autoCombo/provider-family-combos.test.ts timing out at 20s, which reproduces on the PURE release tip under the full vitest suite (and is already tracked by the Release-Green issue #13866), so it is inherited, not this batch's. Merged --admin per merge-gates §3/§4/§7.