Phase 1 / Etch / Concurrency algebra (race, sync, branch, spawn)

[guysenpai]

Closes milestone M1.0.12 — Concurrency algebra (race/sync/branch/spawn { } + cancellation + TaskHandle await). Brief: briefs/M1.0.12-concurrency-algebra.md (Status: CLOSED).

Five review gates (E1–E5), each GO'd on the pushed diff; one Claude.ai STOP round-trip recorded (E0905, §9.2 revision 2) and two in-gate STOP fixes applied (E0905 exemption removal; return-await regression tests).

Closing notes (from the brief)

• What worked: The E1 pointer-stable pool conversion carried the whole milestone — no drive-path pointer ever needed revisiting after it (the brief's "do not attempt an index-based rewrite" note held). Drive-by-origin's index-based pass gave mid-drive child pickup and the deterministic creation-order guarantee for free, and the parent-precedes-children pool-order invariant turned out to be the load-bearing determinism argument (losers are provably canceled before their wakes can fire). The E4/E5 test batteries passed first-run once written against the documented tick timelines.
• What deviated from the original spec: One recorded deviation (STOP round-trip 2026-07-02): the E3 "calls inside the four construct bodies count as consumed launch sites" wording was superseded by etch-resolver-types.md §9.2 revision 2 — await is the SOLE call-grain consumer of the {async} effect; the constructs relocate the suspension, their bodies are ordinary async contexts (E0905 applies recursively). See Recorded deviations.
• What to flag explicitly in review: (1) children_any fires on "any child done OR none suspended" — the second clause realizes the all-branches-failed race passthrough (validated at the E1 GO). (2) Fix-as-you-go: return await <wake-target> silently dropped its return at resume since M1.0.11 — fixed at both resume sites, regression-tested across the three forms (E5 STOP). (3) One-tick construct latency for race/sync with >= 1 admitted branch (parent precedes children in pool order); zero-admitted constructs pass through same-tick — documented in the interleaving tests. (4) Failed tasks park .canceled (shared with explicit cancellation): not a winner, never block a join, await on them fails loud — one state, three semantics, documented on AsyncTask.state. (5) Language-audit exception: one French grammar citation lives in the brief's FROZEN SECTION (upstream-authored, not modifiable here); the two code-comment copies of it were paraphrased to English. (6) KB §9.12 swap content is Claude.ai-produced at the E5 gate (flagged in the gate signal); CLAUDE.md was updated on-branch per §3.4.
• Final measurements: No benchmarks (interpreter milestone, per brief). Test deltas: 934 (base) → 967 tests — +1 E1 substrate, +8 E2 (1 token + 7 parser), +7 E3 types, +7 E4 interp, +9 E5 interp (incl. the 3-form return-await regression) + the §9.6 observable; full suite green debug + ReleaseSafe throughout; zig fmt --check + zig build lint green.
• Residual risks / tech debt left intentionally: (1) Monotonic pool growth — one husk per completed task per program run (brief-accepted; Phase-2 VM has the refcounted lifecycle). (2) Heap-backed values (collections/struct handles) in child-task snapshots and in a race winner's return value share the M0.8 POD-across-suspend rule-arena caveat. (3) A branch block whose LAST item is a bare await parses that await as the block's trailing VALUE → E0904 (M1.0.11 placement, unchanged; the §9.5 pattern ends on return and is unaffected). (4) rules_matched counts once per tick per rule when ANY of its tasks drove — a coarse per-rule notion, informational only. (5) The parent link on race/sync children is lineage bookkeeping only (cancellation stays non-transitive by spec).

Validation checklist (brief step 4)

• Every "Scope" deliverable present (E1 pool substrate; E2 keywords/AST/parse; E3 E0901/E0905/E0906/E0907 + TaskHandle typing; E4 race/sync execution; E5 branch/spawn + runtime TaskHandle)
• No "Out of scope" drift (no bytecode lowering, no transitive cancellation, no timers/wait_unscaled, no entity_event, no value-producing spawn bodies, no pool slot reuse, structural spawn(...) untouched)
• All "Acceptance criteria > Tests" green in debug AND ReleaseSafe (967 tests, 0 failed, 17 pre-existing skips)
• Benchmarks: none required (interpreter milestone)
• Observable behavior demonstrable: §9.5 timeout pattern (winner return propagates) + §9.6 parallel-preload (sync over three awaits) run in the interpreter harness with the documented emit sequences across ticks
• zig build, zig build test (debug + ReleaseSafe), zig fmt --check, zig build lint green
• CLAUDE.md updated on-branch (current-state table, +1 Tags row, M1.0.12 open-decisions entry incl. the §9.2-rev-2 deviation, date)
• §3.6.1 language audit: FR function-word grep on the branch diff = clean (one exception: a French grammar citation inside the brief's FROZEN SECTION, upstream-authored — noted in Closing notes; its two code-comment copies were paraphrased to English)
• §3.6.1 local drift audit: stale pre-M1.0.12 references patched (ast.zig spawn_struct doc, interp.zig AsyncSlot leftover, token/comment sweeps at E2/E3) — no orphans left
• "Closing notes" filled; brief Status: CLOSED, Closed: 2026-07-03
• Final commit docs(brief): close M1.0.12

For Guy: the KB etch-reference-part1.md §9.12 extension (multi-task realization: child tasks, drive-by-origin, WakeCond set, monotonic pool + husk parking, cancellation teardown, one-tick construct latency) is Claude.ai-produced at the E5 gate — to be swapped into the knowledge base alongside the merge + tag (v0.10.12-concurrency-algebra), per brief E5. Merge (squash) and tag are yours.

🤖 Generated with Claude Code