[Epic 2.5] Exact backend: toolexec #349

Summary

The exact-backend spike chose toolexec (RFC-0002 §4.5, decided 2026-10-24). This epic ships it on main. The faultline-toolexec wrapper, run as go test -toolexec, rewrites the test binary's packages so that goroutine starts, channel operations, select, sync primitives, map ranges and math/rand go through sched, a user-space scheduler in the nested module runner/toolexec. The goroutine-mode driver releases one parked goroutine per pass with the seeded toolexec/release stream through internal/gosched, so the same seed gives the same trace hash and the same goroutine order. Runs are labeled exact, replay commands carry the wrapper, FAULTLINE_EXACT=1 refuses a run that would not be exact, and what the rewriter cannot control is listed as coverage gaps. The code moves from spike/exb at d91b8a7 file by file, and the spike's measurements become permanent acceptance tests.

Deliverables

• internal/gosched, the driver hook and XTE-082's stall rule
• Nested module runner/toolexec with the sched runtime and the faultline-toolexec wrapper (prepare, state directory, version line, rewrite set, stale-state message)
• The rewriter with position directives, sorted gaps and the rows of XTE-041 (sync, maps, rand, reflect)
• Package faultline: exact labels, the toolexec replay command, FAULTLINE_EXACT as a bool, toolexec-gaps.txt
• Acceptance package runner/toolexec/acceptance with AT-XTE-26 to AT-XTE-42
• Recorded measurements: point cost, C1 to C3 slowdown, final size

Open questions for the owner

The plan's Task 0 has 12 open questions that wait for the owner, among them exact replay only with Procs 1, FAULTLINE_EXACT as a bool, no trace records for sched decisions, and the size estimate against the 3,000-line limit. Each has a default, and spec XTE is written to the defaults. The refresh task (2.5-00, first in the list below) records the answers, and no code task starts before they are in.

Tasks

• [2.5-00] Refresh plan 2.5, record the owner's answers and apply XTE §12 #335
• [2.5-01] Exact backend: internal/gosched, the driver hook and the watchdog rule #336
• [2.5-02] Exact backend: runner/toolexec module and the sched core #337
• [2.5-03] Exact backend: sched helpers for the new rewrite rows #338
• [2.5-04] Exact backend: wrapper state: prepare, state directory and version line #339
• [2.5-05] Exact backend: wrapper builds: test main, wrapper path, stale state and rewrite set #340
• [2.5-06] Exact backend: rewriter base: positions, gap registry and new points #341
• [2.5-07] Exact backend: rewriter sync rows #342
• [2.5-08] Exact backend: rewriter map, rand and reflect rows; the gap table #343
• [2.5-09] Exact backend: faultline labels, replay command, FAULTLINE_EXACT and gaps #344
• [2.5-10a] Exact backend: end-to-end tests for builds, labels, replay, stalls and gaps #345
• [2.5-10b] Exact backend: end-to-end tests for hangs and orders #346
• [2.5-11] Exact backend: acceptance corpus and the EXB criteria as permanent tests #347
• [2.5-12] Exact backend: measurements and close-out #348

Done when

• Each task's PR is merged into main in task order, one squash commit per task; spike/exb is not merged
• CI's nested job is green on all four runners with the acceptance package (AT-XTE-43)
• The backend's final size is at most 3,000 lines, recorded with each task's line count
• XTE §12's changes are applied (2.5-00) and XTE §9 carries the measured values (2.5-12)

Depends on

• Epic 2.4: Exact backend spike (decision recorded in RFC-0002 §4.5)
• Epics 2.1 to 2.3 merged to main (at or after 4b77273)

References

Private planning docs (.helper/docs/, gitignored).

• Plan: .helper/docs/plans/phase-2/2.5-exact-toolexec.md
• Specs: .helper/docs/specs/exact-toolexec.md, .helper/docs/rfcs/0002-goroutine-determinism.md §4.5, .helper/docs/specs/exact-backend-spike.md, .helper/docs/specs/goroutine-mode.md, .helper/docs/specs/test-api.md