diff --git a/tools/wave/README.md b/tools/wave/README.md
new file mode 100644
index 000000000..ad18714ce
--- /dev/null
+++ b/tools/wave/README.md
@@ -0,0 +1,23 @@
+# tools/wave — the crack-wave harness (P30 S48)
+
+`crack_wave.js` is the CURRENT wave script, kept in the repo because it encodes contracts that were
+paid for in real failures. Invoke with `Workflow({scriptPath: "tools/wave/crack_wave.js", args:[…]})`
+where each arg is `{name, binary, sub, nins, reach, jr, model, prior}`.
+
+**Contracts it encodes — do not drop these when editing:**
+* **Per-agent output dir** `.run/waveN/<fn>/`, and "never delete or modify anything outside your own
+  directory, do not clean up". Wave 2 shared one directory and later agents deleted earlier agents'
+  drafts: 21 adversarially-VERIFIED cracks vanished. (They were all recovered from the run
+  transcripts by `tools/recover_drafts.py` — but only because the transcript is the backup.)
+* **The verifier re-runs `sha1sum` LAST**, after the gate, and reports it separately, so a draft that
+  vanishes mid-run is DETECTED rather than trusted.
+* **Prior-notes seeding.** A target that previously failed is re-run WITH ITS PREDECESSOR'S MEASURED
+  RESIDUAL (`.run/jr48/prior_notes.json`). Measured 7/9 (wave 3) and 10/12 (wave 4). NEAR is a cheap
+  RESUMABLE state, not a write-off.
+* **The cookbook citation list must name every banked block** (§160-§164 today). Idioms are harvested
+  after every wave precisely so the NEXT wave's agents start with them; a stale citation list silently
+  wastes that.
+* **Size-routing** haiku ≤30 ins / sonnet ≤120 / opus above. Wave 2 (20 of 28 Sonnet) beat the
+  all-Opus wave 1 on bank rate.
+
+**Bank rate to date:** 67% → 79% → 69% → 68% of cracks survive the whole-binary gate.
diff --git a/tools/wave/crack_wave.js b/tools/wave/crack_wave.js
new file mode 100644
index 000000000..7fee87b20
--- /dev/null
+++ b/tools/wave/crack_wave.js
@@ -0,0 +1,144 @@
+export const meta = {
+  name: 'p30-s48-stage1-wave4',
+  description: 'Stage-1 wave 4: 39 top-weight families, hardened harness (per-agent dirs, sha1-last verify)',
+  phases: [
+    { title: 'Crack', detail: '39 cracks, size-routed, 9 seeded with prior-attempt notes' },
+    { title: 'Verify', detail: 'independent re-gate + existence proof, sha1 re-checked LAST' },
+  ],
+}
+
+const REPO = '/home/musashi/bfm-decomp'
+const ALL = typeof args === 'string' ? JSON.parse(args) : args
+const TARGETS = ALL.filter(t => t.sub)          // a banked target has no .s — drop it
+
+const CRACK_SCHEMA = {
+  type: 'object',
+  required: ['fn', 'verdict', 'draft_path', 'notes'],
+  properties: {
+    fn: {type: 'string'},
+    verdict: {type: 'string', enum: ['MATCH', 'NEAR', 'FAIL']},
+    draft_path: {type: 'string'},
+    diff_ins: {type: 'integer'},
+    match_one_output: {type: 'string'},
+    sha1: {type: 'string', description: 'sha1sum of the draft you are leaving on disk'},
+    notes: {type: 'string', description: 'idioms, integration surface, exact residual if NEAR; flag any NEW law'},
+  },
+}
+const VERDICT_SCHEMA = {
+  type: 'object',
+  required: ['fn', 'confirmed', 'evidence'],
+  properties: {
+    fn: {type: 'string'},
+    confirmed: {type: 'boolean'},
+    file_exists: {type: 'boolean'},
+    sha1_after: {type: 'string', description: 'sha1sum re-run LAST, after the gate'},
+    evidence: {type: 'string'},
+  },
+}
+
+function crackPrompt(t) {
+  const priorBlock = t.prior ? `
+⚡ A PREVIOUS AGENT ALREADY ATTEMPTED THIS FUNCTION. Read its notes FIRST — they contain the body
+shape, the idioms it proved load-bearing, and the EXACT residual it could not close:
+  cd ${REPO} && python3 -c "import json;print(json.load(open('.run/jr48/prior_notes.json'))['${t.name}'])"
+Treat it as a strong prior from a peer, NOT as gospel: its claims about what is load-bearing are
+reconstructions. If a note contradicts what you measure, the measurement wins — and say so.
+` : ''
+  return `You are cracking ONE function for the Brave Fencer Musashi matching decompilation (PS1, gcc-2.7.2 + maspsx). Repo: ${REPO}.
+
+TARGET: ${t.name} in ${t.binary} — ${t.nins} instructions, family reach ×${t.reach}${t.jr ? ' (mid-function jr / jump table)' : ''}.
+Splat asm subdir: ${t.sub}. Asm file: ${t.sub}/${t.name}.s
+${priorBlock}
+WHY IT MATTERS: zero-crack sibling family — this ONE exemplar templates to ${t.reach - 1} more binaries mechanically.
+
+OUTPUT DISCIPLINE (a previous wave lost 21 verified drafts to a shared directory):
+- Your ONLY output dir is ${REPO}/.run/wave4/${t.name}/ — create it, work there.
+- Final draft MUST be exactly ${REPO}/.run/wave4/${t.name}/${t.name}.c
+- NEVER delete, move or modify ANY file outside your own directory. No \`rm\` outside it, ever.
+  Sibling agents are working in adjacent directories RIGHT NOW.
+- Do NOT clean up at the end. Leave everything in place.
+- Never write to src/, config/, include/ or any tracked file. No make. No git.
+
+METHOD:
+1. STEP 0 — §160g callee-set sibling search (mandatory). Extract the callee set + magic literals from
+   the .s, grep already-matched C in src/ for the same shape. Highest-value hits are IN THE
+   DESTINATION TU ITSELF — an already-matched caller often gives the exact signature, and the host TU
+   fixes every symbol type for free. Read the destination TU before any cross-overlay grep.
+2. A Ghidra seed may exist at .run/ghidra_c/${t.name}.c — VERIFY it is for THIS address before trusting it.
+3. Cookbook — READ THESE FOUR BLOCKS, they are this campaign's own harvest and the most relevant
+   knowledge in the file (~100 entries banked over the last four waves):
+     §160a/b/c and §161a/b/c — the base idioms.
+     §162 — jump-table MAXVAL mirror ("check BOTH edges"), temp SCOPE as a live-range lever, the
+       LICM movable/preheader-order laws, non-void return visible in delay slots, cross-jump
+       DIRECTION as a source-shape oracle, the unreferenced-local frame oracle, and defeating
+       local-alloc's optimize_reg_copy_1 with an in-place shift — a class where REGISTER PINS
+       PROVABLY CANNOT HELP.
+     §163a-e — decl-conflict severity is SCOPE-DEPENDENT (block scope is a conflict SOLVENT); the
+       switch-index parameter-WIDTH oracle (sll/sra straddling the minval subtract);
+       case_values_threshold is 5 (an empty case glued to default can be the only thing that emits a
+       table); cse deleting a copy by rewriting the PREVIOUS insn's SET_DEST (§162j's symptom in a
+       DIFFERENT pass); the frame as a PSEUDO-NUMBER oracle.
+     §164-01..82 — 82 further laws vetted by an independent skeptic each, carrying their evidence
+       grade. **§164z lists 28 REFUTED claims — read it before spending time on a mechanism, so you
+       do not re-derive something already disproved.**
+   Search these before inventing a mechanism: of 190 laws the last waves claimed as new, 57% were
+   already here or did not survive scrutiny.
+4. Iterate: cd ${REPO} && python3 tools/match_one.py ${t.name} --c .run/wave4/${t.name}/${t.name}.c --asm-subdir ${t.sub}
+5. On MATCH: run sha1sum on the draft, report it, and STOP.
+
+RULES: gcc-2.7.2 emits no \`error:\` prefix — read the raw compiler text; "no diagnostic" never means
+"compiler wall". match_one MATCH is a CANDIDATE (§52b) — state your integration surface: which
+callees/globals you declare and whether the host TU declares them differently (§161c). A block-scope
+decl that CONFLICTS with a visible file-scope one is a gate failure; one that AGREES is free — and
+you cannot simply drop the decl, because match_one compiles standalone. Report NEAR honestly with the
+exact residual. Match register allocation exactly (G3).`
+}
+
+function verifyPrompt(c, t) {
+  return `ADVERSARIAL verification, Brave Fencer Musashi decomp. Repo: ${REPO}.
+
+${c.fn} (${t.binary}, ${t.nins} ins) claimed MATCH at ${c.draft_path}; claimed sha1 ${c.sha1 || '(none)'}.
+
+REFUTE it. Default confirmed=false. In this order:
+1. \`ls -l ${c.draft_path}\` and \`sha1sum ${c.draft_path}\` — it must EXIST.
+2. Gate it yourself: cd ${REPO} && python3 tools/match_one.py ${c.fn} --c ${c.draft_path} --asm-subdir ${t.sub}
+   ("FAIL: no C file for <fn>" means the draft is MISSING — that is a refutation, never a pass.)
+3. The draft must DEFINE ${c.fn} (not a renamed/stub body), with no INCLUDE_ASM, no #include of a
+   nonexistent path, no hand-written inline asm standing in for compiled code.
+4. Re-run \`sha1sum ${c.draft_path}\` LAST, after everything else, and report it as sha1_after. If the
+   file vanished or its hash changed since step 1, say so loudly — a previous wave lost 21 verified
+   drafts to exactly that and nothing detected it.
+Do NOT edit or delete anything. Report failures; do not fix them.`
+}
+
+phase('Crack')
+const results = await pipeline(
+  TARGETS,
+  t => agent(crackPrompt(t), {label: `crack:${t.name}`, phase: 'Crack', model: t.model, schema: CRACK_SCHEMA})
+        .then(r => (r ? {...r, t} : null)),
+  r => {
+    if (!r || r.verdict !== 'MATCH' || !r.draft_path) return r
+    return agent(verifyPrompt(r, r.t), {label: `verify:${r.fn}`, phase: 'Verify', schema: VERDICT_SCHEMA})
+      .then(v => ({...r, verdict_check: v}))
+  }
+)
+
+const rows = results.filter(Boolean)
+const confirmed = rows.filter(r => r.verdict === 'MATCH' && r.verdict_check && r.verdict_check.confirmed
+                                   && r.verdict_check.file_exists !== false)
+const refuted = rows.filter(r => r.verdict === 'MATCH' && !confirmed.includes(r))
+const near = rows.filter(r => r.verdict === 'NEAR')
+const fail = rows.filter(r => r.verdict === 'FAIL')
+log(`wave4: confirmed ${confirmed.length} / refuted ${refuted.length} / near ${near.length} / fail ${fail.length} of ${rows.length}`)
+return {
+  counts: {targets: TARGETS.length, returned: rows.length, confirmed: confirmed.length,
+           refuted: refuted.length, near: near.length, fail: fail.length},
+  confirmed: confirmed.map(r => ({fn: r.fn, binary: r.t.binary, reach: r.t.reach, nins: r.t.nins,
+                                  jr: r.t.jr, model: r.t.model, prior: !!r.t.prior,
+                                  draft_path: r.draft_path, sha1: r.sha1, notes: r.notes})),
+  refuted: refuted.map(r => ({fn: r.fn, draft_path: r.draft_path,
+                              evidence: r.verdict_check ? r.verdict_check.evidence : 'verifier died'})),
+  near: near.map(r => ({fn: r.fn, binary: r.t.binary, model: r.t.model, prior: !!r.t.prior,
+                        diff_ins: r.diff_ins, draft_path: r.draft_path, notes: r.notes})),
+  fail: fail.map(r => ({fn: r.fn, binary: r.t.binary, notes: r.notes})),
+}