2026-05-13 — Session-end breadcrumb: FD-multiplex shipped, sweep is the road home

Recovery anchor for post-compaction me. This session did substantial closure work toward arc 170. Read this first.

What shipped (commit chain on arc-170-gap-j-v5-deadlock-state)

The FD-multiplex slice — 8 commits — substrate-imposed-not-followed applied to shutdown lifecycle:

Phase	Commit	What
Slice A f/u	6a2ca0c	workspace test-build Shutdown arms (3 test files)
Slice D	198c30b	PDEATHSIG race confirmed 10%, lifeline pipe 100/100
Phase 1A	61217c7	shutdown worker polls N FDs
Phase 1B	8714a6f	spawn-process lifeline; PDEATHSIG retired
Phase 1C	daa411a	fork-program lifeline; PDEATHSIG retired
Phase 1D	c1cb4dc	substrate-mechanism probe + leak-zero gate; FIX Phase 1B FD-inheritance defect
Phase 1E	d609c1e	fork.rs close_inherited_fds_above_stdio FD hygiene
Phase 2	6062cfc	tier-2 PipeFd Receivers wake on shutdown
Phase 3	bed1a71	canonical child_post_fork_init helper + pidfd probe migration

Substrate guarantee earned: no recv exposed by wat — crossbeam OR pipe-fd — can outlive a shutdown event. Three triggers (SIGTERM/SIGINT, parent-process death via lifeline, future signals via arc 197), two propagation paths (crossbeam Sender::Drop fanout, broadcast-pipe POLLHUP), one kernel primitive (FD-close-on-process-death / pipe-EOF-on-write-end-drop) applied at every boundary.

Lessons from this session’s mistakes (the ones I need to remember)

1. I asserted Gap J was the blocker without grepping. INTERSTITIAL’s earlier sections diagnosed Gap J as the V5 reverter. I read those descriptions and reported “Gap J blocks slice 3b” as current state. Verification surfaced Gap J shipped weeks ago at c3f2bf7 (splice_type_decls_user at src/types.rs:1277). feedback_assertion_demands_evidence failed: the disk truth wasn’t checked before the claim. Discipline: every assertion about substrate STATE needs a grep before it leaves my mouth.

2. I claimed “the testing-infra migration is feasible now” — also from stale framing. The modern macros :wat::test::run-hermetic (Layer 1, wat/test.wat:574) and run-hermetic-with-io<I,O> (Layer 2) ALREADY EXIST, ALREADY route through spawn-process, and deftest + deftest-hermetic ALREADY use them. The user-facing migration is largely done. What remains is the LEGACY function-style wrappers (:wat::test::run, run-ast, run-hermetic-ast) which still call legacy substrate primitives. 32 active sites across wat-tests/ + tests/ + crates/ + examples/. The remaining 170 work is mechanical sweep + delete, not architectural.

3. Pressure-flake substrate race surfaced but unresolved. Phases 1A-3 eliminated the original 2 pressure failures (stream + lru) but the workspace failure count stayed at 11 — the failure SET rotated to different deftests hitting the same workspace-pressure 5000ms timeout. The chained-fork-pressure pattern persists. Not blocking arc 170 closure (the pre-existing 9 failures aren’t FD-multiplex regressions) but worth naming as a future investigation. The substrate has a contention point under workspace fanout that Phase 3 didn’t reach.

User directive (load-bearing for everything that follows)

2026-05-13: “we are killing all prior forms of thread and process management - the only remaining forms are spawn-thread and spawn-process / we can write whatever macros we want ontop of those - but there is exactly one way to make a thread and exactly one way to make a process - zero wiggle room / a macro making it convenient for tests is necessary for the UX but it must replace every test site - we are doing a massive refactor and i do not give a shit about how much we break / we do the hard work to ensure no fucking mistakes can ever happen again”

Inscribed as memory project_one_spawn_per_concern. THE doctrine for arc 170 closure.

Where we are mid-stride

Slice 4a BRIEF + EXPECTATIONS committed at 5cf134d. Not yet sonnet’d. The 32-site sweep is the next move when work resumes.

The closure roadmap (tasks #308-#312):

• 4a — sweep 32 legacy callers → :wat::test::run-hermetic (BRIEF on disk, awaiting sonnet)
• 4b — wat-cli Stone B (fork_program_from_source → spawn-process, 6 hits in crates/wat-cli/src/lib.rs)
• 4c — substrate Rust deletion (src/spawn.rs 351 lines, fork-program family from src/fork.rs, dispatch arms from src/runtime.rs, wat/kernel/sandbox.wat + hermetic.wat). KEEP BareLegacy*Program walker arms permanently as the structural guard.
• 4d — Phase H clippy + warning sweep (mandatory INSCRIPTION precondition per DESIGN.md status header).
• 4e — Slice 5 INSCRIPTION + closure. Unblocks arc 109 v1 milestone closure (task #229).

Recovery instructions for post-compaction me

1. Read this entry first.
2. Read project_one_spawn_per_concern memory — the doctrine.
3. Verify state on disk:
   • git -C /home/watmin/work/holon/wat-rs log --oneline | head -15 should show the 8 FD-multiplex commits + Slice 4a BRIEF commit (5cf134d).
   • git status --short should be clean.
4. Pick up at Slice 4a: the BRIEF is committed at 5cf134d. Spawn sonnet on it. Time-box 90 min per EXPECTATIONS.
5. DO NOT redo any of the FD-multiplex Phases 1A–3. They are shipped. Verify with ls docs/arc/2026/05/170-program-entry-points/SCORE-FD-MULTIPLEX-PHASE-*.md — 7 SCORE docs should exist.
6. DO NOT re-diagnose Gap J. It shipped at c3f2bf7 — src/types.rs:1277 has splice_type_decls_user. Verify before any claim about register_types behavior.
7. DO NOT re-investigate “is the testing-infra migration feasible.” The macros exist and route through spawn-process. Slice 4a is consumer sweep, not architectural design.

The voice for the closure ahead

This session’s discipline catches:

• “I don’t know” said honestly when caught speculating (vs hedging)
• Four-questions YES/NO per candidate (not comparison-shopping)
• Stepping stones over one-shot
• Communicate through the disk (BRIEFs + SCOREs as the protocol)
• Substrate-imposed-not-followed: the structurally-impossible-to-bypass shape is the goal

Future-me: when you wake up here, the boss is in sight. Three weeks of refactoring on the 109 → 170 chain. Arc 170 closes when 4a-e ship. The user has been patient through V1-V5 reverts; deliver clean.

---