clojurewasm · chaploud · Jul 16, 2026 · Jul 16, 2026
diff --git a/.dev/debt.yaml b/.dev/debt.yaml
@@ -3,30 +3,6 @@
 # Refresh on every /continue resume — .claude/skills/continue/RESUME.md Step 0.5.
 
 entries:
-  - id: "D-521"
-    layer: "code"
-    status: "note"
-    description: |-
-      Binary-size campaign lever #2 (ADR-0204 D2): complete the table-driven
-      arm64 dispatch and DELETE the legacy switch. `engine.codegen.arm64.emit
-      .compile` is a single 707 KB __text symbol. Root cause sharpened
-      2026-07-16: `dispatch_collector.dispatch()` is an `inline for` +
-      `@call(.auto)` over the 409 registered arm64 op modules — every
-      registered handler body is INLINED into the compile() instantiation
-      (and dispatch is a linear O(n) compare chain in the JIT-compile hot
-      path), on top of the ~161 legacy switch arms inlining their op_*.zig
-      helpers. Stage A = convert dispatch()/dispatchX86_64Ctx() to a
-      comptime-built fn-pointer table indexed by @intFromEnum(op) (O(1)
-      dispatch, kills the inline duplication; handlers need a uniform
-      canonical signature per axis — wrap where inferred error sets
-      differ). Stage B = migrate the remaining ~161 switch ops to per-op
-      files and delete the switch. Acceptance: emitted JIT bytes
-      IDENTICAL — test-aot-diff 63/63 + fuzz-diff lanes + 3-OS test-all.
-    first_raised: "2026-07-16"
-    last_reviewed: "2026-07-16"
-    refs: |-
-      ADR-0204 D2; from_cljw_05 request 1; dispatch_collector ADR-0074
-    front: binary-size
   - id: "D-522"
     layer: "api"
     status: "note"
@@ -41,12 +17,15 @@ entries:
       **Stage 1 DONE 2026-07-16** (shared noinline fpBridge1/fpBridge2
       bodies; thunks forward-only): jit_host_bridge 1,311 KB -> 232 KB
       (-82%), whole CLI 5,282,584 -> 4,173,736 B (-21%), no ABI change.
-      Stage 2 (collapse the slot axis by conveying K at runtime — call
-      site materializes the slot index before BLR) is RE-SCORED by the
-      stage-1 measurement: remaining recoverable ~200 KB (209 KB of
-      per-slot thunks -> ~15 KB shared), against a JIT call-sequence
-      change on both arches + AOT format touch. Demand-driven — re-score
-      after D-521 lands; do not grind speculatively (measure-first).
+      Campaign CLOSED with stage 1 (v2.2.1; ADR-0204 Implemented).
+      RESIDUAL = stage 2 only: collapse the x64 slot axis by conveying K
+      at runtime (call site materializes the slot index before BLR) —
+      remaining recoverable ~200 KB (209 KB of per-slot thunks -> ~15 KB
+      shared), against a JIT call-sequence change on both arches + AOT
+      format touch. Demand-driven; trigger = a real embedder squeezed on
+      size after v2.2.1, or the next size campaign. Do not grind
+      speculatively (measure-first; cf. the D-521 refutation lesson
+      2026-07-16-outlining-once-called-handlers-size-neutral.md).
       Constraints unchanged: no api break, no per-call regression beyond
       noise (ADR-0181 bench series).
     first_raised: "2026-07-16"

diff --git a/.dev/decisions/0204_binary_size_campaign.md b/.dev/decisions/0204_binary_size_campaign.md
@@ -1,6 +1,10 @@
 # ADR-0204: Binary-size campaign — host-bridge thunk collapse + table-driven emitter completion
 
-- **Status**: Accepted (kickoff; stages land as separate PRs)
+- **Status**: Implemented (D1 stage 1 shipped in #145: jit_host_bridge
+  −82%, CLI −21% ReleaseSafe / −8% ReleaseFast; D1 stage 2 re-scored
+  ~200 KB → demand-driven in D-522; D2 size premise refuted by stage-A
+  measurement → D-521 discharged, see 2026-07-16 amendment below; D3 not
+  pursued. Released as v2.2.1.)
 - **Date**: 2026-07-16
 - **Front**: binary-size (D-521 / D-522; dogfooding request `from_cljw_05`)
 - **Findings base**: symbol-attribution measurement 2026-07-16 (below);
@@ -72,6 +76,16 @@ t1fp = 3×5×64 = 960, t0..t4 = 5×5×64 = 1,600. Two independent findings:
 
 ### D2 — Complete the table-driven arm64 dispatch; delete the legacy switch (D-521, lever #2: ~500 KB potential)
 
+> **Amended 2026-07-16 — size premise REFUTED by stage-A measurement.**
+> Converting `dispatch()` to a comptime fn-pointer table left
+> `emit.compile` at 707 KB and grew the binary +28.8 KB (reverted). The
+> 707 KB symbol is *aggregation* of once-called inlined handlers, not
+> duplication — out-lining is size-neutral. See lesson
+> `2026-07-16-outlining-once-called-handlers-size-neutral.md`. D-521 is
+> discharged; the per-op-file migration of the remaining ~161 switch arms
+> stays on the ADR-0074 maintainability trajectory, NOT as a size lever.
+> The original text below is retained for the record.
+
 `engine.codegen.arm64.emit.compile` (src/engine/codegen/arm64/emit.zig,
 ~2,100-line fn) is a ~161-arm switch over `ZirOp` whose arms mostly
 delegate to already-extracted `op_*.zig` handlers. The table-driven

diff --git a/.dev/handover.md b/.dev/handover.md
@@ -11,26 +11,27 @@ PIC codegen D-516 fix; release.yml auto-built Release + 5 assets; release was
 USER-DIRECTED 2026-07-09 per ADR-0156). v1 frozen at `v1.11.1`. Dev model: cut
 a `develop/<slug>` branch from `main` → PR → CI `ci-required` 3-OS gate must be
 green to merge. **Release stays user-only (ADR-0156)** — never autonomously tag /
-publish / cut over. Active campaign: binary-size (ADR-0204, below); no cron
-self-re-arm.
-
-## Active front — binary-size campaign (ADR-0204, kickoff 2026-07-16)
-
-Trigger = dogfooding mailbox `from_cljw_05` (cljw measures zwasm at 44% of its
-shipped code; user authorized an independent zwasm-side campaign). Measured
-baseline (ReleaseSafe arm64 CLI 5,282,584 B @71cccba76; ReleaseFast base
-DOUBLED since 2026-06-12: 1,972,696 → 3,884,856 B):
-- **Lever #1 = D-522**: `api.jit_host_bridge` comptime thunk cross-product =
-  1,311 KB / 5,453 syms (t2fp×2,880 + t1fp×960 + t0..4). Characterize the
-  instantiation axes first, then collapse to descriptor-driven trampolines.
-  ABI/api unchanged; bench-gated.
-- **Lever #2 = D-521**: `engine.codegen.arm64.emit.compile` = one 707 KB
-  symbol (~161-arm ZirOp switch). Finish dispatch_collector (ADR-0074)
-  registration, delete the fall-through switch; x86_64 twin same treatment.
-  Acceptance: emitted JIT bytes IDENTICAL (test-aot-diff + fuzz-diff).
-- Re-run `record_binary_size.sh` per stage-merge (rows into the same PR).
-- cljw reply posted (`to_cljw_05.md`): thunk cross-product ≫ api-surface
-  gating (ADR-0073 already covers the latter); x86_64 emitter already exists.
+publish / cut over (v2.2.1 tag was user-granted 2026-07-16 in-session). No
+active campaign; no cron self-re-arm.
+
+## Binary-size campaign — CLOSED 2026-07-16 (ADR-0204 Implemented, v2.2.1)
+
+Trigger = dogfooding mailbox `from_cljw_05` (cljw measures zwasm at 44% of
+its shipped code). Kickoff #144 (measured attribution re-prioritized the
+cljw asks) · stage 1 #145 (D-522: shared noinline fpBridge1/2 bodies —
+jit_host_bridge 1,311→232 KB, **CLI −21%** ReleaseSafe, −8% ReleaseFast;
+DA-critique 20/20) · close #146 (refutation record + v2.2.1). Key outcomes:
+- **D-522 stage 1 SHIPPED**; stage 2 (slot-axis, ~200 KB) re-scored →
+  demand-driven note in debt.yaml.
+- **D-521 DISCHARGED — size premise refuted by measurement**: fn-ptr-table
+  stage A left `emit.compile` at 707 KB, +28.8 KB binary → reverted. The
+  giant symbol is once-called-handler AGGREGATION, not duplication. Lesson
+  `2026-07-16-outlining-once-called-handlers-size-neutral.md` (predictive
+  question = "how many call sites share this code?", not symbol size).
+- ReleaseFast `base` had DOUBLED unnoticed since 2026-06-12 (1.97→3.88 MB;
+  series cadence = phase-boundary only); now 3.56 MB. cljw replies
+  `to_cljw_05/06.md` in the mailbox (x86_64 emitter already exists +
+  4.0 MB budget-line revisit suggestion).
 
 ## AOT-full-fidelity campaign — CLOSED 2026-07-09 (ADR-0203 Implemented)
 
@@ -105,7 +106,7 @@ Report = `.dev/meta_audits/2026-07-06-senior-runtime-gap-analysis.md`.
   dogfooded into cljw (pins zwasm by git tag-hash). Runners ReleaseSafe.
 - **EH**: cross-instance JIT EH both arches. Interp+JIT EH corpus green. Realworld 56
   fixtures interp 56/0; JIT diff-gated.
-- **Debt**: 68 entries — **0 `now`-class** (D-505 DONE: 3 arm64-SIMD bitmask sites
+- **Debt**: 69 entries — **0 `now`-class** (D-505 DONE: 3 arm64-SIMD bitmask sites
   spill-aware, bitselect/fma SPILL-EXEMPT, spill_aware promoted to CI+gate_commit;
   follow-on D-506 = FP spill stage-2, note-class). 完成形 plateau (all dims
   confirmed, surface audits clean, interp+JIT fuzz 0-crash, v1-JIT parity D-265 closed).

diff --git a/.dev/lessons/2026-07-16-outlining-once-called-handlers-size-neutral.md b/.dev/lessons/2026-07-16-outlining-once-called-handlers-size-neutral.md
@@ -0,0 +1,44 @@
+# Out-lining once-called inlined handlers is size-NEUTRAL (a giant symbol ≠ duplication)
+
+- **Date**: 2026-07-16
+- **Area**: engine/codegen dispatch; binary size (ADR-0204 campaign)
+- **Trigger**: D-521 stage A — converting `dispatch_collector.dispatch()`
+  from `inline for` + `@call(.auto)` to a comptime fn-pointer table,
+  expecting `engine.codegen.arm64.emit.compile` (one 707 KB `__text`
+  symbol) to shrink by the out-lined handler share.
+
+## Observation
+
+Measured (ReleaseSafe arm64 CLI, on top of D-522 stage 1):
+
+- `emit.compile`: 707,492 B → 707,540 B — **unchanged**.
+- Whole binary: 4,173,736 → 4,202,536 B — **+28.8 KB** (+213 handler
+  symbols; part of the delta is symtab/strtab, not code).
+- Change reverted.
+
+## Why (re-derivable)
+
+A function inlined into its **single** call site is not duplicated —
+the code exists exactly once, merely aggregated under the caller's
+symbol. Out-lining it moves the same bytes behind a `call` and ADDS
+call overhead + a symbol-table row. Per-symbol size attribution
+(`nm` next-addr − addr) makes such a caller look like a monolith to
+attack, but symbol size ≠ recoverable size. Contrast D-522 stage 1
+(same campaign, same day): there the bodies were monomorphized into
+**64 copies each** (per-slot comptime K) — genuine duplication — and
+sharing them recovered −1.08 MB. **The question that predicts the
+outcome is "how many call sites / instantiations share this code?",
+not "how big is the symbol?"**
+
+Same family as the D-507 scalar-elision retrospective ("biggest
+lever" refuted by measurement): size/perf hypotheses from symbol
+attribution or peer-project claims (here from_cljw_05's "table-driven
+encoder typically shrinks 50–80%") must be probed with the cheapest
+reversible experiment before scheduling a migration campaign.
+
+## Residual
+
+The per-op-file migration of the remaining ~161 legacy switch arms
+(ADR-0074 trajectory) and O(1) table dispatch remain *maintainability/
+compile-speed* options — pursue them on their own merits if ever
+measured to matter, never as a size lever (D-521 discharged).
diff --git a/.dev/lessons/INDEX.md b/.dev/lessons/INDEX.md
@@ -24,6 +24,7 @@
 
 | Date       | Slug                                                     | Keywords                                                                                                                                                                                                                                                                                                                                                                                                                    | One-line                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            |                                                                                                                                                                                                                                                                                                                                                                                                      |
 |------------|----------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| 2026-07-16 | outlining-once-called-handlers-size-neutral | binary size, symbol size not recoverable size, giant emit.compile 707KB aggregation not duplication, inline for dispatch fn-pointer table stage A refuted +28.8KB reverted, once-called inline = no duplication, D-522 per-slot 64-copies genuine duplication -1.08MB, how many call sites share this code, probe cheapest reversible experiment before migration campaign, from_cljw_05 50-80% claim refuted, D-521 discharged | D-521 stage A (fn-ptr table dispatch replacing inline-for) measured: `emit.compile` UNCHANGED at 707 KB, binary +28.8 KB → reverted. A once-called inlined handler is aggregated, not duplicated — out-lining is size-neutral minus overhead. Contrast D-522 stage 1 (−1.08 MB): per-slot comptime K monomorphized each body ×64 = genuine duplication. Predictive question = "how many call sites/instantiations share this code?", not "how big is the symbol?". Probe before scheduling migration campaigns. |  |
 | 2026-07-09 | failure-path-tests-certified-the-defect | ADR-0203 D5 cache silent-miss, failure-path testing, unit test asserted defective behavior, DA critique reproduced corrupt-entry brick + ZWASM_DEBUG fail + interp override, happy-path bias, tests mirror implementation not invariant, enumerate failure modes red-first, self-heal degrade-not-fail | Stage-5 `--cache` shipped the happy path complete while the D5 invariant "cache can never make run fail" was implemented only for read errors; the unit test ASSERTED the defect (corrupt entry returned as hit) and a same-file comment contradicted the module doc. DA critique reproduced 4 failing-run shapes. Lesson: enumerate an invariant's failure modes and write one red test per mode BEFORE implementing — implementer-written tests otherwise mirror the implementation, not the invariant. Fixed 4a17ceda7 (header-gate + self-heal, refusal=bypass, interp bypass). |  |
 | 2026-07-09 | stale-cache-binary-false-differential-finding | stale .zig-cache/o/<hash> exe path, per-edit new hash dir old binary survives, fault-injection artifact mimics real finding, MEM-MISMATCH offset 0x3 +1 signature, re-resolve via zig build step, disambiguate by running candidates on committed known-green corpora, clean-rebuild green before trusting post-injection findings | Running a differential runner by a remembered `.zig-cache/o/<hash>/` path after source edits executed a STALE (fault-injected) build — its lone campaign "finding" was the injection artifact. Each edit compiles to a NEW hash dir; old ones survive. Rules: re-resolve the exe via `zig build <step>` after any edit (or verify the candidate against the committed known-green corpora first); end every fault-injection experiment with a clean rebuild + green run; a finding mirroring a recent deliberate perturbation is presumed contamination until a fresh build reproduces it. |  |
 | 2026-07-01 | listen-test-runner-endofstream-guest-stdout | failed command --listen=-, phantom test failure exit 0, Zig 0.16 test runner EndOfStream panic after all pass, WASI guest stdout real fd 1 corrupts protocol, builtin.is_test guard fd.zig signal.zig | `zig build test` shows a phantom `failed command: … --listen=-` even when all tests pass — a WASI test wrote guest stdout to real fd 1 (the --listen protocol channel); guard real std-fd writes with `!builtin.is_test` |