feat(bughunt): single-finder hunt with graph-fed targeting, telemetry, and defect fixes (#1364)

* chore: ignore graphify-out * fix(bughunt): assemble the report in-script - the report agent dropped findings * fix(bughunt): retry only unverified candidates and catch garbage-verdict batches * fix(bughunt): retune the round budget floor and require a budget directive * feat(bughunt): per-round telemetry and a runStats aggregate Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(bughunt-run): record run telemetry and validate coordinates after each hunt * feat(bughunt): drop the sonnet finder slot Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(bughunt): rebuild adaptive targeting - directory clusters, cooldown, graph-fed explore lenses * fix(bughunt): rewind a dead finder's explore draw so unread files are never marked clean * docs(bughunt-run): pre-hunt graph ranking checklist and offline test roster * fix(bughunt): rewind thrown-stage explore draws and correct log/doc wording Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(bughunt): disambiguate absorb() drop log and retire dead sonnet label alternation Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
2026-09-03 03:50:00 +03:00 · 2026-08-13 16:52:23 +02:00
parent 3af3489f71
commit 34f2e41207
4 changed files with 585 additions and 240 deletions
@@ -15,14 +15,29 @@ reach a commit, an issue, or a PR body.

 ## 1. Hunt

-Read the ledger and pass every record in as `known`, so the hunt does not
-re-derive anything already found, fixed, declined, or refuted:
+**Launch the hunt from a turn that carries a token-budget directive** (recommended:
+`+25M`, comfortably above a full 8-round run). The workflow's cost ceiling is gated
+on `budget.total`, which is null without a directive — a directive-less run has **no
+ceiling at all**. The workflow's first log line echoes the state: `budget=25M` means
+armed; `budget=NONE - cost ceiling disarmed` means stop the run and relaunch with a
+directive.
+
+Before launching, in order:
+
+1. **Rebuild the graph** (stale coordinates aim the explore lens at moved code):
+   `graphify update . --no-cluster` — local tree-sitter, zero LLM cost, ~10.7k nodes.
+2. **Rank explore targets**: `node .superpowers/rank-explore.mjs` — writes
+   `.superpowers/explore-ranking.json`, deprioritizing files recorded clean in
+   `.superpowers/explored-clean.json` and dropping files that no longer exist.
+3. Read the ledger and pass every record in as `known`, so the hunt does not
+   re-derive anything already found, fixed, declined, or refuted.

 ```
 Workflow({
  name: "bughunt",
  args: {
    known: <every record from findings-ledger.json, as {file, line, title, status}>,
+    graph: <the rows of .superpowers/explore-ranking.json>,
    lenses: [ {key, prompt}, ... ],   // optional: scope the hunt to one subsystem
    maxRounds: 8,
    dryThreshold: 2,
@@ -30,6 +45,10 @@ Workflow({
 })
 ```

+If `graph` is omitted or empty the hunt logs
+`explore: args.graph absent/empty - falling back to churn-based fresh eyes` and
+still runs — degraded targeting, never a smaller lens family.
+
 Omit `lenses` for a general hunt across the rotating families.

 Each lens object is `{key, prompt}`. `key` must match `^[a-z0-9-]+$` —
@@ -40,7 +59,21 @@ validated — it reaches the finder prompt as the literal string `undefined`,
 silently degrading that lens instead of failing loudly. Check your lens
 objects before passing them.

-When it returns, append each entry of `result.confirmed` to the ledger with
+When it returns, first save the raw result verbatim to
+`.superpowers/hunts/<YYYY-MM-DD>-raw.json`, then:
+
+```bash
+node .superpowers/render-run-stats.mjs .superpowers/hunts/<YYYY-MM-DD>-raw.json <hunt-name>
+```
+
+This validates the result shape, appends the run's telemetry to
+`.superpowers/run-history.json`, updates `.superpowers/explored-clean.json`, and
+checks **every** confirmed finding's coordinates against the working tree (file
+exists, line within length — the report agent that used to spot-check two findings
+is gone). Resolve any `COORD` warnings before appending to the ledger: stale
+coordinates poison `bughunt-fix`.
+
+Then append each entry of `result.confirmed` to the ledger with
 `status: "open"`, an id from `nextId`, and today's date. Bump `nextId`. The
 incoming record carries a prose `fix` field (bughunt's suggested remedy) —
 rename it to `suggestedFix` when appending, so the ledger's `fix` field starts
@@ -51,15 +84,12 @@ revertProof}` once something is actually fixed. Then:
 node .superpowers/render-ledger.mjs
 ```

-Each confirmed record carries `finder` — which model in the dual-model panel
-produced it. The run also logs one `panel:` line with the split. The two finders
-are unioned, not voted, so the second model's entire value is what it finds
-alone; because duplicates collapse to the opus-slot record, a `sonnet` tag means
-opus missed it. Watch that count across a few hunts. Consistently zero is the
-evidence for dropping to a single finder — but note that would also weaken
-convergence, since a round is only allowed to count as dry when the full panel
-reported, so a lone finder having a bad day would read as "clean" instead of
-"we didn't fully look".
+Each confirmed record carries `finder: "opus"`. The dual-model finder panel was
+retired 2026-08-12: attribution over the only measured run priced sonnet's unique
+yield (1 high, 4 medium, 9 low) at roughly a third of the run's agents. The known
+cost: with one finder, a lazy-but-non-null finder round can read as "clean" where
+the panel required both models to agree it was. Watch `runStats` — per-lens
+candidate counts make an anomalously empty lens visible after the fact.

 ## 2. Gate (human)

@@ -159,7 +189,9 @@ node .claude/workflows/bughunt.harness.mjs
 node .claude/workflows/bughunt-fix.harness.mjs
 node .superpowers/render-ledger.mjs --selftest
 node .superpowers/verify-fixes.mjs --selftest
+node .superpowers/rank-explore.mjs --selftest
+node .superpowers/render-run-stats.mjs --selftest
 ```

-All four run offline with zero API calls. Run them after any edit to the
+All six run offline with zero API calls. Run them after any edit to the
 relevant script.
@@ -43,18 +43,17 @@ export async function run({ agentStub, args = undefined, budget = undefined }) {
 }

 // ---------- stub kit (used from Task 2 onward; harmless now) ----------
-export function makeStub({ hunt, verify, report = () => 'REPORT_MD', recon = defaultRecon }) {
+export function makeStub({ hunt, verify, recon = defaultRecon }) {
  return (prompt, opts) => {
    const label = opts.label || ''
    if (label.startsWith('recon:')) return recon(label)
-    let m = /^r(\d+):hunt:([a-z0-9-]+):(opus|sonnet)$/.exec(label)
-    if (m) return hunt(Number(m[1]), m[2], m[3], prompt)
+    let m = /^r(\d+):hunt:([a-z0-9-]+):opus$/.exec(label)
+    if (m) return hunt(Number(m[1]), m[2], 'opus', prompt)
    m = /^r(\d+):verify:([a-z0-9-]+?)(:retry)?$/.exec(label)
    if (m) {
      const candidates = JSON.parse(prompt.split('--- CANDIDATES ---')[1])
      return verify(Number(m[1]), m[2], candidates, Boolean(m[3]), prompt)
    }
-    if (label === 'report') return report(prompt)
    throw new Error(`unexpected agent label: ${label}`)
  }
 }
@@ -72,6 +71,8 @@ export const finding = (n, over = {}) => ({
  evidence: 'e',
  ...over,
 })
+export const graphRows = (n) =>
+  Array.from({ length: n }, (_, i) => ({ file: `Server/gen/g${i}.go`, score: 1 - i / (n + 1), degree: 10, cited: 5 }))
 export const confirmAll = (cands) => ({
  verdicts: cands.map((c) => ({
    title: c.title, file: c.file, line: c.line,
@@ -92,16 +93,11 @@ const scenarios = {}

 // S1: happy convergence - one bug in round 1, rounds 2-3 dry -> converged.
 scenarios.s1_convergence = async () => {
-  const reportPrompts = []
-  const { result, calls } = await run({
+  const { result, calls, logs } = await run({
    agentStub: makeStub({
      hunt: (round, key, model) =>
        round === 1 && key === 'ws-hub' && model === 'opus' ? { findings: [finding(1)] } : none,
      verify: (round, key, cands) => confirmAll(cands),
-      report: (prompt) => {
-        reportPrompts.push(prompt)
-        return 'REPORT_MD'
-      },
    }),
  })
  for (const k of ['converged', 'stoppedOnBudget', 'rounds', 'confirmed', 'unverified', 'report'])
@@ -111,23 +107,23 @@ scenarios.s1_convergence = async () => {
  assert.deepEqual(result.rounds.map((r) => r.dryAfter), [0, 1, 2])
  assert.deepEqual(result.rounds.map((r) => r.family), ['surfaces', 'bug-classes', 'flows'])
  assert.equal(result.confirmed.length, 1)
-  assert.equal(result.report, 'REPORT_MD')
  assert.ok(!calls.some((c) => (c.opts.label || '').startsWith('r4:')), 'no round 4 after convergence')
-  assert.match(reportPrompts[0], /CONVERGED after 3 round\(s\)/)
-  assert.match(reportPrompts[0], /\| 1 \| surfaces \|/)
+  assert.ok(!calls.some((c) => c.opts.label === 'report'), 'the report is built in-script')
+  assert.match(result.report, /CONVERGED after 3 round\(s\)/)
+  assert.match(result.report, /### high - distinct bug alpha1 omega1/)
+  assert.match(result.report, /\| 1 \| surfaces \|/)
+  assert.ok(logs.some((l) => /budget=NONE - cost ceiling disarmed/.test(l)), 'a directive-less run must announce the dead ceiling')
 }

-// S2: panel dedupe - opus and sonnet report the same bug -> one candidate, one verify call.
-scenarios.s2_panel_dedupe = async () => {
+// S2: near-duplicate findings from a single finder collapse - one candidate, one verify call.
+scenarios.s2_finder_dedupe = async () => {
  const verifyBatches = []
-  const { result } = await run({
+  const { result, calls } = await run({
    agentStub: makeStub({
-      hunt: (round, key, model) => {
-        if (round !== 1 || key !== 'ws-hub') return none
-        return model === 'opus'
-          ? { findings: [finding(1, { line: 100 })] }
-          : { findings: [finding(1, { line: 105, title: 'distinct bug alpha1 omega1 variant' })] }
-      },
+      hunt: (round, key) =>
+        round === 1 && key === 'ws-hub'
+          ? { findings: [finding(1, { line: 100 }), finding(1, { line: 105, title: 'distinct bug alpha1 omega1 variant' })] }
+          : none,
      verify: (round, key, cands) => {
        verifyBatches.push(cands)
        return confirmAll(cands)
@@ -137,6 +133,9 @@ scenarios.s2_panel_dedupe = async () => {
  assert.equal(verifyBatches.length, 1)
  assert.equal(verifyBatches[0].length, 1)
  assert.equal(result.confirmed.length, 1)
+  const hunts = calls.filter((c) => /^r1:hunt:ws-hub:/.test(c.opts.label || ''))
+  assert.equal(hunts.length, 1, 'exactly one finder call per lens - the sonnet slot is gone')
+  assert.match(hunts[0].opts.label, /:opus$/, 'the label keeps the :opus suffix the harness parses')
 }

 // S3: refuted findings stay dead - re-reported next round, never re-verified; refutes count toward dry.
@@ -237,15 +236,132 @@ scenarios.s6b_verifier_double_failure = async () => {
  assert.equal(result.unverified.length, 0, 'later-confirmed candidate must leave the unverified list')
 }

-// S7: rounds 1-3 each confirm a bug -> round 4 runs adaptive lenses built from the stats.
+// N1 (spec #1): the top-ranked hotspot cluster sits out exactly the next round, then returns.
+// The producing cluster keeps running when eligible (the old s7b lock, restated under cooldown).
+scenarios.s_cluster_cooldown = async () => {
+  const { calls } = await run({
+    args: { maxRounds: 6, dryThreshold: 9, graph: graphRows(60) },
+    agentStub: makeStub({
+      hunt: (round, key) => {
+        if (round === 1 && key === 'ws-hub')
+          return { findings: [
+            finding(1, { file: 'Server/ws/hub.go', title: 'ws bug alpha one' }),
+            finding(2, { file: 'Server/ws/pubsub.go', line: 300, title: 'ws bug beta two' }),
+          ] }
+        if (round === 2 && key === 'concurrency')
+          return { findings: [
+            finding(3, { file: 'Server/ws/emit.go', title: 'ws bug gamma three' }),
+            finding(4, { file: 'Server/api/user.go', title: 'api bug delta four' }),
+          ] }
+        if (round === 4 && key === 'hotspot-server-ws')
+          return { findings: [finding(9, { file: 'Server/ws/late.go', title: 'late ws bug nine' })] }
+        return none
+      },
+      verify: (round, key, cands) => confirmAll(cands),
+    }),
+  })
+  const hunted = (rnd, key) => calls.some((c) => (c.opts.label || '') === `r${rnd}:hunt:${key}:opus`)
+  assert.ok(hunted(4, 'hotspot-server-ws'), 'top cluster hunts in r4')
+  assert.ok(!hunted(5, 'hotspot-server-ws'), 'the r4 top cluster must sit out r5 even though it produced')
+  assert.ok(hunted(5, 'hotspot-server-api'), 'the next cluster takes the top slot in r5')
+  assert.ok(hunted(6, 'hotspot-server-ws'), 'cooldown lasts exactly one round')
+}
+
+// N2 (spec #2): a cooldown gap FREEZES cleanStreak - neither increments nor resets - so
+// demotion still means two consecutive clean APPEARANCES. If the gap incremented, ws would
+// be demoted before r6; if demotion broke, ws would still run in r8.
+scenarios.s_cooldown_freezes_streak = async () => {
+  const { result, calls } = await run({
+    args: { maxRounds: 8, dryThreshold: 9, graph: graphRows(100) },
+    agentStub: makeStub({
+      hunt: (round, key) => {
+        if (round === 1 && key === 'ws-hub')
+          return { findings: [
+            finding(1, { file: 'Server/ws/hub.go', title: 'ws bug alpha one' }),
+            finding(2, { file: 'Server/ws/pubsub.go', line: 300, title: 'ws bug beta two' }),
+          ] }
+        if (round === 2 && key === 'concurrency')
+          return { findings: [finding(3, { file: 'Server/api/user.go', title: 'api bug delta three' })] }
+        return none
+      },
+      verify: (round, key, cands) => confirmAll(cands),
+    }),
+  })
+  const hunted = (rnd, key) => calls.some((c) => (c.opts.label || '') === `r${rnd}:hunt:${key}:opus`)
+  assert.ok(hunted(4, 'hotspot-server-ws'), 'clean appearance #1 in r4')
+  assert.ok(!hunted(5, 'hotspot-server-ws'), 'cooldown in r5')
+  assert.ok(hunted(6, 'hotspot-server-ws'), 'the gap must freeze the streak at 1, not increment it')
+  assert.equal(result.rounds.length, 8, 'the run must reach r8 for the demotion assert to mean anything')
+  assert.ok(!hunted(8, 'hotspot-server-ws'), 'two clean appearances (r4, r6) demote the lens')
+}
+
+// N3 (spec #3): cooldown+demotion emptying the hotspot pool must backfill from explore and
+// log it - silent family shrinkage is the exact freshEyesLens() defect this rebuild removes.
+scenarios.s_hotspot_backfill = async () => {
+  const { calls, logs } = await run({
+    args: { maxRounds: 5, dryThreshold: 9, graph: graphRows(100) },
+    agentStub: makeStub({
+      hunt: (round, key) =>
+        round === 1 && key === 'ws-hub'
+          ? { findings: [finding(1, { file: 'Server/ws/hub.go', title: 'lone ws bug one' })] }
+          : none,
+      verify: (round, key, cands) => confirmAll(cands),
+    }),
+  })
+  const r5 = calls.filter((c) => /^r5:hunt:/.test(c.opts.label || '')).map((c) => c.opts.label.split(':')[2])
+  assert.ok(!r5.some((k) => k.startsWith('hotspot-')), 'the sole cluster is on cooldown in r5')
+  assert.deepEqual([...r5].sort(), ['explore-1', 'explore-2', 'explore-3', 'explore-4'], 'the family backfills to full size from explore')
+  assert.ok(logs.some((l) => /hotspot pool short/.test(l)), 'backfill must be logged, never silent')
+}
+
+// N4 (spec #4): within-run consumption - later rounds draw the NEXT chunk of the ranking,
+// never re-offering files already handed to an explore lens this run.
+scenarios.s_explore_consumption = async () => {
+  const { calls } = await run({
+    args: { maxRounds: 5, dryThreshold: 9, graph: graphRows(80) },
+    agentStub: makeStub({
+      hunt: (round, key) =>
+        round === 1 && key === 'ws-hub'
+          ? { findings: [finding(1, { file: 'Server/ws/hub.go', title: 'seed bug one' })] }
+          : none,
+      verify: (round, key, cands) => confirmAll(cands),
+    }),
+  })
+  const promptOf = (rnd, key) => (calls.find((c) => (c.opts.label || '') === `r${rnd}:hunt:${key}:opus`) || {}).prompt || ''
+  assert.match(promptOf(4, 'explore-1'), /Server\/gen\/g0\.go/)
+  assert.match(promptOf(4, 'explore-2'), /Server\/gen\/g10\.go/)
+  assert.match(promptOf(4, 'explore-3'), /Server\/gen\/g20\.go/, 'r4 backfills a third explore lens (single cluster)')
+  assert.match(promptOf(5, 'explore-1'), /Server\/gen\/g30\.go/, 'r5 draws the next chunk')
+  assert.doesNotMatch(promptOf(5, 'explore-1'), /Server\/gen\/g0\.go/, 'r5 must not re-offer r4 files')
+}
+
+// N6 (spec #6): args.graph absent -> churn-based fresh-eyes fallback, logged, family intact.
+scenarios.s_graph_missing_fallback = async () => {
+  const { calls, logs } = await run({
+    args: { maxRounds: 4, dryThreshold: 9 },
+    agentStub: makeStub({
+      hunt: (round, key) =>
+        round === 1 && key === 'ws-hub'
+          ? { findings: [finding(1, { file: 'Server/ws/hub.go', title: 'seed bug one' })] }
+          : none,
+      verify: (round, key, cands) => confirmAll(cands),
+    }),
+  })
+  assert.ok(logs.some((l) => /falling back to churn/.test(l)), 'the fallback must be logged')
+  const e1 = calls.find((c) => (c.opts.label || '') === 'r4:hunt:explore-1:opus')
+  assert.ok(e1, 'an explore lens must still run from churn')
+  assert.match(e1.prompt, /Server\/api\/user\.go/, 'churned file with no findings feeds the fallback')
+}
+
+// S7: rounds 1-3 each confirm a bug -> round 4 runs adaptive lenses: directory-granularity
+// hotspots plus explore (churn fallback here - no args.graph is passed).
 scenarios.s7_adaptive_lenses = async () => {
  const A = finding(1, { file: 'Server/ws/hub.go', line: 120, title: 'alpha race window one' })
  const B = finding(2, { file: 'Server/ws/pubsub.go', line: 60, title: 'beta subscription leak two' })
  const C = finding(3, { file: 'Client/tauri-client/src/lib/livekitE2EE.ts', line: 200, title: 'gamma epoch desync three' })
  const { result, calls } = await run({
    agentStub: makeStub({
-      hunt: (round, key, model) => {
-        if (model !== 'opus') return none
+      hunt: (round, key) => {
        if (round === 1 && key === 'ws-hub') return { findings: [A] }
        if (round === 2 && key === 'concurrency') return { findings: [B] }
        if (round === 3 && key === 'flow-voice') return { findings: [C] }
@@ -257,49 +373,26 @@ scenarios.s7_adaptive_lenses = async () => {
  assert.equal(result.converged, true)
  assert.equal(result.rounds.length, 5) // r4, r5 adaptive + dry
  assert.equal(result.rounds[3].family, 'adaptive')
-  const r4Hunts = calls.filter((c) => /^r4:hunt:/.test(c.opts.label || ''))
-  const r4Keys = [...new Set(r4Hunts.map((c) => c.opts.label.split(':')[2]))]
+  const r4Keys = [...new Set(calls.filter((c) => /^r4:hunt:/.test(c.opts.label || '')).map((c) => c.opts.label.split(':')[2]))]
  assert.ok(r4Keys.includes('hotspot-server-ws'), `r4 keys: ${r4Keys}`)
-  assert.ok(r4Keys.includes('fresh-eyes'), `r4 keys: ${r4Keys}`)
-  const hotspot = r4Hunts.find((c) => c.opts.label.includes('hotspot-server-ws'))
+  assert.ok(r4Keys.includes('hotspot-client-tauri-client-src-lib'), `r4 keys: ${r4Keys}`)
+  assert.ok(r4Keys.includes('explore-1'), `r4 keys: ${r4Keys}`)
+  const hotspot = calls.find((c) => (c.opts.label || '').includes('hotspot-server-ws'))
  assert.match(hotspot.prompt, /Server\/ws\/hub\.go/)
  assert.match(hotspot.prompt, /alpha race window one/)
-  const freshEyes = r4Hunts.find((c) => c.opts.label.includes('fresh-eyes'))
-  assert.match(freshEyes.prompt, /Server\/api\/user\.go/) // churned, never a finding
+  const explore = calls.find((c) => (c.opts.label || '') === 'r4:hunt:explore-1:opus')
+  assert.match(explore.prompt, /Server\/api\/user\.go/) // churned, never a finding
  assert.equal(result.confirmed.length, 3)
 }

-// S7b: a lens with 2 consecutive clean rounds is demoted from later rounds.
-scenarios.s7b_demotion = async () => {
-  const { result, calls } = await run({
-    args: { maxRounds: 6 },
-    agentStub: makeStub({
-      hunt: (round, key, model) => {
-        if (model !== 'opus') return none
-        const src = { 1: 'ws-hub', 2: 'concurrency', 3: 'flow-reconnect', 4: 'hotspot-server-ws', 5: 'hotspot-server-ws' }
-        if (key === src[round])
-          return { findings: [finding(round, { file: `Server/ws/a${round}.go`, title: `unique bug number${round} zeta${round}` })] }
-        return none
-      },
-      verify: (round, key, cands) => confirmAll(cands),
-    }),
-  })
-  const labels = calls.map((c) => c.opts.label || '')
-  assert.ok(labels.some((l) => /^r5:hunt:fresh-eyes:/.test(l)), 'fresh-eyes still runs in r5 (streak 1)')
-  assert.ok(!labels.some((l) => /^r6:hunt:fresh-eyes:/.test(l)), 'fresh-eyes demoted in r6 (streak 2)')
-  assert.ok(labels.some((l) => /^r6:hunt:hotspot-server-ws:/.test(l)), 'producing hotspot keeps running')
-  assert.equal(result.converged, false)
-  assert.equal(result.confirmed.length, 5)
-}
-
-// S7c: a lens whose VERIFIER died is not demoted; a zero-candidate lens still is.
+// S7c: a lens whose VERIFIER died is not demoted (its cluster returns after cooldown);
+// a zero-candidate explore lens still accrues streak and demotes.
 scenarios.s7c_verifier_failure_not_demoted = async () => {
  const early = { 1: 'ws-hub', 2: 'concurrency', 3: 'flow-reconnect' }
  const { result, calls } = await run({
-    args: { maxRounds: 6 },
+    args: { maxRounds: 6, dryThreshold: 9, graph: graphRows(100) },
    agentStub: makeStub({
-      hunt: (round, key, model) => {
-        if (model !== 'opus') return none
+      hunt: (round, key) => {
        if (round <= 3 && key === early[round])
          return { findings: [finding(round, { file: `Server/ws/a${round}.go`, title: `early bug item${round} kappa${round}` })] }
        if (round >= 4 && key === 'hotspot-server-ws')
@@ -310,12 +403,15 @@ scenarios.s7c_verifier_failure_not_demoted = async () => {
    }),
  })
  const labels = calls.map((c) => c.opts.label || '')
-  assert.ok(labels.some((l) => /^r6:hunt:hotspot-server-ws:/.test(l)), 'verifier-dead lens must NOT be demoted')
-  assert.ok(!labels.some((l) => /^r6:hunt:fresh-eyes:/.test(l)), 'zero-candidate lens still accrues streak and demotes')
+  assert.ok(labels.some((l) => /^r4:hunt:hotspot-server-ws:/.test(l)))
+  assert.ok(!labels.some((l) => /^r5:hunt:hotspot-server-ws:/.test(l)), 'cooldown after topping r4')
+  assert.ok(labels.some((l) => /^r6:hunt:hotspot-server-ws:/.test(l)), 'a verifier-dead lens must NOT be demoted')
+  assert.ok(!labels.some((l) => /^r6:hunt:explore-1:/.test(l)), 'a zero-candidate explore lens still demotes')
  assert.equal(result.confirmed.length, 3)
-  assert.equal(result.unverified.length, 3)
+  assert.equal(result.unverified.length, 2) // b4 and b6, each denied a verdict twice
  assert.equal(result.converged, false)
-  assert.ok(result.rounds.slice(3).every((r) => r.dryEligible === false))
+  assert.equal(result.rounds[3].dryEligible, false)
+  assert.equal(result.rounds[5].dryEligible, false)
 }

 // S12: empty adaptive family (no confirms, no churn) must break honestly, not count dry rounds.
@@ -354,7 +450,7 @@ scenarios.s8_budget_floor = async () => {
 scenarios.s8b_budget_midrun = async () => {
  let n = 0
  const { result } = await run({
-    budget: { total: 1000000, spent: () => 0, remaining: () => (n++ === 0 ? 200000 : 100000) },
+    budget: { total: 10000000, spent: () => 0, remaining: () => (n++ === 0 ? 3000000 : 1000000) },
    agentStub: makeStub({
      hunt: (round, key, model) =>
        round === 1 && key === 'ws-hub' && model === 'opus' ? { findings: [finding(1)] } : none,
@@ -496,7 +592,7 @@ scenarios.s_custom_lenses = async () => {
    agentStub: makeStub({ hunt: () => none, verify: (r, k, c) => confirmAll(c) }),
  })
  const keys = calls
-    .map((c) => /^r1:hunt:([a-z0-9-]+):(opus|sonnet)$/.exec(c.opts.label || ''))
+    .map((c) => /^r1:hunt:([a-z0-9-]+):opus$/.exec(c.opts.label || ''))
    .filter(Boolean)
    .map((m) => m[1])
  assert.deepEqual([...new Set(keys)].sort(), ['voice-e2ee-keyholder', 'voice-e2ee-rotation'])
@@ -549,50 +645,7 @@ scenarios.s_confirmed_carries_finder_detail = async () => {
  assert.equal(r.fix, 'FIX_TEXT')
  assert.equal(r.lens, 'ws-hub')
  assert.equal(r.round, 1)
-}
-
-// S_FINDER_ATTRIBUTION: confirmed records name which model found them. The panel unions rather
-// than votes, so this is the only way to tell whether the second finder earns its cost. Because
-// dedupe keeps the first occurrence and opus is index 0, `finder: 'sonnet'` means opus did NOT
-// report it - i.e. a sonnet-unique find. `finder: 'opus'` says nothing about sonnet either way.
-scenarios.s_finder_attribution = async () => {
-  const shared = finding(1)
-  // Different file, not just a different line: isDup treats same-file findings within
-  // TITLE_MATCH_WINDOW as duplicates on title-word overlap, and the finding() fixtures share
-  // enough words to collapse into one another.
-  const sonnetOnly = finding(2, { file: 'Server/api/user.go' })
-  const { result } = await run({
-    args: { maxRounds: 1, dryThreshold: 9 },
-    agentStub: makeStub({
-      hunt: (round, key, model) => {
-        if (round !== 1 || key !== 'ws-hub') return none
-        return model === 'opus' ? { findings: [shared] } : { findings: [shared, sonnetOnly] }
-      },
-      verify: (round, key, cands) => confirmAll(cands),
-    }),
-  })
-  const byTitle = Object.fromEntries(result.confirmed.map((r) => [r.title, r.finder]))
-  assert.equal(byTitle[shared.title], 'opus', 'a find both models made keeps the opus-first record')
-  assert.equal(byTitle[sonnetOnly.title], 'sonnet', 'a find only sonnet made must be attributed to sonnet')
-}
-
-// S_FINDER_ATTRIBUTION_SURVIVES_DEAD_OPUS: the tag must be taken from the panel slot, not from the
-// position in the surviving list. Filtering the nulls out BEFORE reading the index shifts sonnet
-// into slot 0 and mislabels every one of its finds as opus - exactly when attribution matters most.
-scenarios.s_finder_attribution_survives_dead_opus = async () => {
-  const only = finding(3)
-  const { result } = await run({
-    args: { maxRounds: 1, dryThreshold: 9 },
-    agentStub: makeStub({
-      hunt: (round, key, model) => {
-        if (round !== 1 || key !== 'ws-hub') return none
-        return model === 'opus' ? null : { findings: [only] }
-      },
-      verify: (round, key, cands) => confirmAll(cands),
-    }),
-  })
-  assert.equal(result.confirmed.length, 1)
-  assert.equal(result.confirmed[0].finder, 'sonnet', 'a dead opus must not relabel sonnet finds as opus')
+  assert.equal(r.finder, 'opus', 'the finder tag is constant now but the ledger still expects it')
 }

 // S_VERIFIER_IS_NOT_TOLD_THE_FINDER: the verifier prompt deliberately says "another model" and
@@ -619,26 +672,162 @@ scenarios.s_verifier_is_not_told_the_finder = async () => {
  }
 }

-// S_PANEL_SPLIT_IS_REPORTED: a tag nobody reads is not a measurement. The run must say out loud
-// how many confirmed findings only the second finder produced, which is the number that decides
-// whether the second model is worth its cost.
-scenarios.s_panel_split_is_reported = async () => {
-  const { logs } = await run({
+// New (spec Testing #8): the report is built in-script. Section count must equal the confirmed
+// count at 82 (the agent version emitted 79 for 82), and the unverified section must survive
+// the agent's removal - it used to exist only inside the report agent's prompt.
+scenarios.s_report_deterministic = async () => {
+  const many = Array.from({ length: 82 }, (_, i) =>
+    finding(i, { file: `Server/ws/f${i}.go`, line: 10, title: `unique bug row${i} tag${i}` }))
+  const stuck = finding(999, { file: 'Server/api/stuck.go', line: 40, title: 'stuck bug never verified' })
+  const { result, calls } = await run({
    args: { maxRounds: 1, dryThreshold: 9 },
+    agentStub: makeStub({
+      hunt: (round, key, model) =>
+        round === 1 && key === 'ws-hub' && model === 'opus' ? { findings: [...many, stuck] } : none,
+      verify: (round, key, cands) => confirmAll(cands.filter((c) => c.file !== 'Server/api/stuck.go')),
+    }),
+  })
+  assert.equal(result.confirmed.length, 82)
+  assert.equal(result.unverified.length, 1)
+  assert.ok(!calls.some((c) => c.opts.label === 'report'), 'no report agent may run')
+  const sections = (result.report.match(/^### /gm) || []).length
+  assert.equal(sections, 82, 'one section per confirmed finding, none dropped')
+  assert.match(result.report, /## Unverified - re-run/)
+  assert.match(result.report, /stuck bug never verified/)
+  assert.match(result.report, /## Convergence/)
+}
+
+// New (spec Testing #7): the retry re-sends ONLY unmatched candidates, and N garbage verdicts
+// (count == candidate count, zero of them matching) must still trigger it - the hole S10 misses
+// because S10's verdict list is empty rather than full of junk.
+scenarios.s_targeted_retry = async () => {
+  const a = finding(1, { file: 'Server/ws/a.go', title: 'alpha bug one paired' })
+  const b = finding(2, { file: 'Server/api/b.go', title: 'beta bug two orphaned' })
+  const retryBatches = []
+  const { result } = await run({
+    args: { maxRounds: 1, dryThreshold: 9 },
+    agentStub: makeStub({
+      hunt: (round, key, model) =>
+        round === 1 && key === 'ws-hub' && model === 'opus' ? { findings: [a, b] } : none,
+      verify: (round, key, cands, isRetry) => {
+        if (isRetry) {
+          retryBatches.push(cands)
+          return confirmAll(cands)
+        }
+        // one real verdict for a, one garbage verdict pointing nowhere: count matches, content doesn't
+        return {
+          verdicts: [
+            { title: a.title, file: a.file, line: a.line, refuted: false, reason: 'ok', confidence: 'high', severity: 'high', fix: 'f' },
+            { title: 'hallucinated', file: 'Server/nowhere.go', line: 1, refuted: false, reason: 'x', confidence: 'low', severity: 'low' },
+          ],
+        }
+      },
+    }),
+  })
+  assert.equal(retryBatches.length, 1, 'retry must fire despite verdict count == candidate count')
+  assert.deepEqual(retryBatches[0].map((c) => c.file), ['Server/api/b.go'], 'only the unmatched candidate is re-sent')
+  assert.equal(result.confirmed.length, 2)
+  assert.equal(result.unverified.length, 0)
+}
+
+// New (spec Testing #9): the retuned floor must stop a run the old 150k floor let through.
+// 1M remaining is under the ~2M measured per-round cost - starting a round would overshoot.
+scenarios.s9_budget_ceiling_retuned = async () => {
+  const { result, logs } = await run({
+    budget: { total: 10000000, spent: () => 9000000, remaining: () => 1000000 },
+    agentStub: makeStub({ hunt: () => none, verify: (r, k, c) => confirmAll(c) }),
+  })
+  assert.equal(result.rounds.length, 0, '1M remaining must not start a ~2M round')
+  assert.equal(result.stoppedOnBudget, true)
+  assert.ok(logs.some((l) => /Budget floor/.test(l)))
+}
+
+// New: telemetry. Per-round suppression split (ledger vs same-run), spend sampling, file
+// coverage, severity mix, per-lens precision, and the top-level runStats aggregate. Without
+// this every cost figure from a run is eyewitness-only - the 2026-08-12 problem.
+scenarios.s_telemetry = async () => {
+  let spent = 0
+  const known = [{ file: 'Server/ws/hub.go', line: 100, title: 'known bug from ledger prior', status: 'fixed' }]
+  const { result } = await run({
+    args: { maxRounds: 1, dryThreshold: 9, known },
+    budget: { total: 50000000, spent: () => (spent += 500000), remaining: () => 40000000 },
    agentStub: makeStub({
      hunt: (round, key, model) => {
-        if (round !== 1 || key !== 'ws-hub') return none
-        return model === 'opus'
-          ? { findings: [finding(1)] }
-          : { findings: [finding(1), finding(2, { file: 'Server/api/user.go' })] }
+        if (round !== 1 || key !== 'ws-hub' || model !== 'opus') return none
+        return { findings: [
+          finding(1, { file: 'Server/ws/hub.go', line: 102, title: 'known bug from ledger prior' }),
+          finding(2, { file: 'Server/api/fresh.go', severity: 'medium', title: 'fresh bug beta gamma' }),
+        ] }
      },
      verify: (round, key, cands) => confirmAll(cands),
    }),
  })
-  const line = logs.find((l) => /panel:/.test(l))
-  assert.ok(line, 'the run must report the panel split')
-  assert.match(line, /sonnet-only/, 'the split must name the sonnet-only count explicitly')
-  assert.match(line, /\b1\b/, 'exactly one confirmed finding here was sonnet-only')
+  const r1 = result.rounds[0]
+  assert.equal(r1.suppressedLedger, 1, 'the ledger-known duplicate must be counted as ledger suppression')
+  assert.equal(r1.suppressedRun, 0)
+  assert.ok(r1.spentAfter > r1.spentBefore, 'per-round spend must be sampled')
+  assert.equal(r1.filesTouched, 1)
+  assert.equal(r1.filesNew, 1)
+  assert.deepEqual(r1.severity, { critical: 0, high: 0, medium: 1, low: 0 })
+  assert.equal(r1.perLens['ws-hub'].confirmed, 1)
+  assert.equal(r1.perLens['ws-hub'].fresh, 1)
+  assert.ok(result.runStats, 'runStats missing from the result')
+  assert.equal(result.runStats.confirmed, 1)
+  assert.equal(result.runStats.suppressedLedger, 1)
+  assert.equal(result.runStats.config.maxRounds, 1)
+  assert.match(result.report, /## Run stats/)
+}
+
+// N7 (Task 9 review finding): a dead finder on an explore lens read nothing - its draw is
+// rewound so the files never reach exploredFiles, where the session would record them clean
+// and deprioritize them in every future hunt. maxRounds caps at 4 on purpose: a live round-5
+// lens would legitimately re-read the rewound files and they would CORRECTLY re-enter
+// exploredFiles - the poison-prevention property is only assertable when the run ends here.
+// Re-offering in later rounds follows from the same exploreConsumed state drawExploreFiles
+// filters on, so this one scenario locks the mechanism.
+scenarios.s_explore_rewind_on_dead_finder = async () => {
+  const { result, calls } = await run({
+    args: { maxRounds: 4, dryThreshold: 9, graph: graphRows(80) },
+    agentStub: makeStub({
+      hunt: (round, key) => {
+        if (round === 1 && key === 'ws-hub')
+          return { findings: [finding(1, { file: 'Server/ws/hub.go', title: 'seed bug one' })] }
+        if (round === 4 && key === 'explore-1') return null // dead finder: read nothing
+        return none
+      },
+      verify: (round, key, cands) => confirmAll(cands),
+    }),
+  })
+  const promptOf = (rnd, key) => (calls.find((c) => (c.opts.label || '') === `r${rnd}:hunt:${key}:opus`) || {}).prompt || ''
+  assert.match(promptOf(4, 'explore-1'), /Server\/gen\/g0\.go/, 'r4 explore-1 drew the head of the ranking')
+  for (let i = 0; i < 10; i++)
+    assert.ok(!result.exploredFiles.includes(`Server/gen/g${i}.go`), `g${i} was never read - must not be reported explored`)
+  assert.ok(result.exploredFiles.includes('Server/gen/g10.go'), 'files a LIVE lens drew stay reported')
+  assert.ok(result.exploredFiles.includes('Server/gen/g20.go'), 'backfilled live lens files stay reported too')
+}
+
+// N8 (final-review finding): a THROWN stage nulls the whole lens result - the second
+// finder-failure mode the code documents. Its explore draw must rewind exactly like the
+// null-finder case, or never-read files reach exploredFiles and poison explored-clean.
+scenarios.s_explore_rewind_on_thrown_stage = async () => {
+  const { result, calls } = await run({
+    args: { maxRounds: 4, dryThreshold: 9, graph: graphRows(80) },
+    agentStub: makeStub({
+      hunt: (round, key) => {
+        if (round === 1 && key === 'ws-hub')
+          return { findings: [finding(1, { file: 'Server/ws/hub.go', title: 'seed bug one' })] }
+        if (round === 4 && key === 'explore-1') throw new Error('finder infrastructure blew up')
+        return none
+      },
+      verify: (round, key, cands) => confirmAll(cands),
+    }),
+  })
+  const promptOf = (rnd, key) => (calls.find((c) => (c.opts.label || '') === `r${rnd}:hunt:${key}:opus`) || {}).prompt || ''
+  assert.match(promptOf(4, 'explore-1'), /Server\/gen\/g0\.go/, 'r4 explore-1 drew the head of the ranking')
+  for (let i = 0; i < 10; i++)
+    assert.ok(!result.exploredFiles.includes(`Server/gen/g${i}.go`), `g${i} was never read - must not be reported explored`)
+  assert.ok(result.exploredFiles.includes('Server/gen/g10.go'), 'files a LIVE lens drew stay reported')
+  assert.equal(result.rounds[3].dryEligible, false, 'a nulled lens result still makes the round ineligible')
 }

 // ---------- runner ----------
@@ -1,10 +1,9 @@
 export const meta = {
  name: 'bughunt',
-  description: 'Converging multi-round bug hunt: rotating lens families, dual-model panels, fable refute-by-default verification, dry-threshold stop',
+  description: 'Converging multi-round bug hunt: rotating lens families, single opus finder, fable refute-by-default verification, dry-threshold stop',
  whenToUse: 'Hunting real bugs across the Go server, Tauri Rust backend, and TS client until consecutive rounds go dry. Not a security-only scan.',
  phases: [
    { title: 'Recon', detail: 'haiku: churn + concurrency-surface inventory' },
-    { title: 'Report', detail: 'fable: ranked findings + convergence table' },
  ],
 }

@@ -19,14 +18,15 @@ const ARGS = (() => {
 const MAX_ROUNDS = ARGS.maxRounds || 8
 const DRY_THRESHOLD = ARGS.dryThreshold || 2
 // A scoped hunt (args.lenses) replaces the round-1 family outright; later rounds still go
-// adaptive, so hotspot and fresh-eyes coverage - and therefore convergence - still work.
+// adaptive, so hotspot and explore coverage - and therefore convergence - still work.
 const CUSTOM_LENSES = Array.isArray(ARGS.lenses) && ARGS.lenses.length ? ARGS.lenses : null
-// ponytail: rough floor for one round (up to 12 high-effort finders + verifiers); tune after live runs
-const ROUND_BUDGET_FLOOR = 150000
+// Floor for one round, tuned from the 2026-08-12 run: ~2.6M output tokens per round measured.
+// The old 150k floor would overshoot the ceiling by nearly a full round.
+const ROUND_BUDGET_FLOOR = 2000000
 // The args channel has already been observed delivering something the script
 // could not read; an unnoticed fallback here is an 8x cost surprise, so say out
 // loud what the run is actually going to do.
-log(`config: maxRounds=${MAX_ROUNDS} dryThreshold=${DRY_THRESHOLD}${CUSTOM_LENSES ? ` lenses=custom(${CUSTOM_LENSES.length})` : ''}`)
+log(`config: maxRounds=${MAX_ROUNDS} dryThreshold=${DRY_THRESHOLD}${CUSTOM_LENSES ? ` lenses=custom(${CUSTOM_LENSES.length})` : ''} budget=${budget.total ? Math.round(budget.total / 1e6) + 'M' : 'NONE - cost ceiling disarmed'}`)

 // ---------- schemas: copied VERBATIM from the current bughunt.js ----------
 const FINDINGS = {
@@ -93,14 +93,16 @@ Out of scope, do not report: naming, formatting, missing tests, "consider adding
 performance that is not a hang, anything you cannot point at specific lines for.

 Method:
-  1. Read the actual files. Never report from a filename or a grep hit alone.
+  1. Read the actual files. Never report from a filename, a grep hit, or a graph edge alone - a
+     graphify edge is structural evidence of coupling, not of a bug; open the cited file and confirm.
  2. For every candidate, grep for ALL callers before judging - a guard may already live upstream.
  3. Check whether an existing test already locks the behavior you think is wrong. If a test asserts it,
     it is intended behavior, not a bug. Test files are *_test.go and tests/unit/*.test.ts.
  4. Report EVERY finding you can prove - there is no cap. The quality bar stays: zero findings is a
     valid, respectable answer, and each finding needs file, line, and a concrete repro.

-You may run read-only shell commands (grep, git log, go doc). Do not modify any file. Do not run the test suite.
+You may run read-only shell commands (grep, git log, go doc, graphify path, graphify explain).
+Do not modify any file. Do not run the test suite.
 `

 // ---------- lens catalog ----------
@@ -271,46 +273,83 @@ const FLOW_LENSES = [
 ]

 function lensesForRound(round) {
-  if (CUSTOM_LENSES) return round === 1 ? CUSTOM_LENSES : buildAdaptiveLenses()
+  if (CUSTOM_LENSES) return round === 1 ? CUSTOM_LENSES : buildAdaptiveLenses(round)
  if (round === 1) return SURFACE_LENSES
  if (round === 2) return BUGCLASS_LENSES
  if (round === 3) return FLOW_LENSES
-  return buildAdaptiveLenses()
+  return buildAdaptiveLenses(round)
 }
 function familyName(round) {
  if (CUSTOM_LENSES) return round === 1 ? 'custom' : 'adaptive'
  return ['surfaces', 'bug-classes', 'flows'][round - 1] || 'adaptive'
 }
+// Directory granularity: the old two/three-segment cluster collapsed the whole TS client into
+// one bucket (35 of 82 findings), so the "top cluster" never changed for five straight rounds.
 function clusterOf(file) {
  const parts = String(file).split('/')
-  return parts.slice(0, parts[0] === 'Client' ? 3 : 2).join('/')
+  return parts.length > 1 ? parts.slice(0, -1).join('/') : parts[0]
 }
-function freshEyesLens() {
-  const files = churnFiles.filter((f) => !seen.some((s) => s.file === f)).slice(0, 10)
-  if (!files.length) return []
-  return [
-    {
-      key: 'fresh-eyes',
-      prompt:
-        `These files churned heavily in the last 8 weeks, yet no hunt round has confirmed or refuted a ` +
-        `single finding in them - either they are clean or every lens so far walked past them. Read each ` +
-        `one IN FULL with fresh eyes and hunt for real bugs of any class:\n` +
-        files.map((f) => `  - ${f}`).join('\n'),
-    },
-  ]
+
+// ---------- explore targeting ----------
+// args.graph: session-computed coupling ranking (rank-explore.mjs). The workflow only reads
+// .file - scoring already happened outside, where the filesystem is.
+const GRAPH_ROWS = (Array.isArray(ARGS.graph) ? ARGS.graph : []).filter((r) => r && typeof r.file === 'string')
+const EXPLORE_FILES_PER_LENS = 10
+const exploreConsumed = new Set() // within-run consumption: never re-offer a file to a later round
+let exploreFallbackLogged = false
+function drawExploreFiles() {
+  let pool
+  if (GRAPH_ROWS.length) pool = GRAPH_ROWS.map((r) => r.file)
+  else {
+    if (!exploreFallbackLogged) {
+      log('explore: args.graph absent/empty - falling back to churn-based fresh eyes')
+      exploreFallbackLogged = true
+    }
+    pool = churnFiles
+  }
+  const files = pool
+    .filter((f) => !exploreConsumed.has(f) && !seen.some((s) => s.file === f))
+    .slice(0, EXPLORE_FILES_PER_LENS)
+  for (const f of files) exploreConsumed.add(f)
+  return files
 }
-function buildAdaptiveLenses() {
+function exploreLens(i) {
+  const files = drawExploreFiles()
+  if (!files.length) return null
+  const src = GRAPH_ROWS.length
+    ? `These files are heavily coupled (per the code graph) to files where confirmed bugs live, yet no ` +
+      `hunt has confirmed or refuted a single finding in them - either they are clean or every lens so ` +
+      `far walked past them.`
+    : `These files churned heavily in the last 8 weeks, yet no hunt round has confirmed or refuted a ` +
+      `single finding in them - either they are clean or every lens so far walked past them.`
+  return {
+    key: `explore-${i}`,
+    prompt: `${src} Read each one IN FULL with fresh eyes and hunt for real bugs of any class:\n` +
+      files.map((f) => `  - ${f}`).join('\n'),
+    files,
+  }
+}
+
+let cooldownCluster = null // the top-ranked cluster hunted in round N sits out round N+1
+function buildAdaptiveLenses(round) {
  const byCluster = {}
  for (const c of confirmedAll) {
    const cl = clusterOf(c.file)
    if (!byCluster[cl]) byCluster[cl] = []
    byCluster[cl].push(c)
  }
-  const top = Object.entries(byCluster)
+  // Explore-heavy schedule: measured hotspot yield flattened to 0.25 high+med/agent by round 6.
+  const hotspotQuota = round <= 5 ? 2 : 1
+  const exploreQuota = round <= 5 ? 2 : 3
+  const hotKey = (cl) => ('hotspot ' + cl).toLowerCase().replace(/[^a-z0-9]+/g, '-')
+  const picked = Object.entries(byCluster)
    .sort((a, b) => b[1].length - a[1].length)
-    .slice(0, 3)
-  const hotspots = top.map(([cluster, items]) => ({
-    key: ('hotspot ' + cluster).toLowerCase().replace(/[^a-z0-9]+/g, '-'),
+    .filter(([cl]) => cl !== cooldownCluster)
+    .filter(([cl]) => (cleanStreak[hotKey(cl)] || 0) < 2) // pre-filter so backfill sees the real shortfall
+    .slice(0, hotspotQuota)
+  cooldownCluster = picked.length ? picked[0][0] : null
+  const hotspots = picked.map(([cluster, items]) => ({
+    key: hotKey(cluster),
    prompt:
      `Bugs cluster. Confirmed findings so far in ${cluster}:\n` +
      items.map((i) => `  - ${i.file}:${i.line} ${i.title}`).join('\n') +
@@ -318,7 +357,19 @@ function buildAdaptiveLenses() {
      `(subscribe/unsubscribe, open/close, register/transfer, acquire/release), and the paths a past fix ` +
      `here did NOT cover. Do not re-report the findings listed above - they are already known.`,
  }))
-  return [...hotspots, ...freshEyesLens()]
+  const shortfall = hotspotQuota - hotspots.length
+  if (shortfall > 0) log(`adaptive: hotspot pool short by ${shortfall} - trying explore backfill`)
+  const explores = []
+  for (let i = 1; i <= exploreQuota + shortfall; i++) {
+    if ((cleanStreak[`explore-${i}`] || 0) >= 2) continue // demoted slot: no substitution, that IS demotion
+    const lens = exploreLens(i)
+    if (!lens) {
+      log(`adaptive: explore pool exhausted after ${explores.length} lens(es)`)
+      break
+    }
+    explores.push(lens)
+  }
+  return [...hotspots, ...explores]
 }

 // ---------- dedupe + ledger helpers ----------
@@ -343,10 +394,18 @@ function isDup(a, b) {
  const hits = aw.filter((w) => bw.has(w)).length
  return hits * 2 >= aw.length
 }
-function dedupe(cands, priors) {
+function dedupe(cands, priors, counts) {
  const kept = []
  for (const c of cands) {
-    if (priors.some((p) => isDup(c, p)) || kept.some((k) => isDup(c, k))) continue
+    const prior = priors.find((p) => isDup(c, p))
+    if (prior) {
+      if (counts) counts[prior.fromLedger ? 'suppressedLedger' : 'suppressedRun']++
+      continue
+    }
+    if (kept.some((k) => isDup(c, k))) {
+      if (counts) counts.suppressedRun++
+      continue
+    }
    kept.push(c)
  }
  return kept
@@ -417,6 +476,7 @@ const seen = (ARGS.known || []).map((k) => ({
  line: k.line,
  title: k.title,
  status: k.status || 'known',
+  fromLedger: true, // telemetry: distinguishes ledger suppression from same-run suppression
 }))
 const confirmedAll = []
 const unverified = []
@@ -464,10 +524,12 @@ while (dry < DRY_THRESHOLD && round < MAX_ROUNDS) {
  const family = lensesForRound(round + 1)
  if (!family || !family.length) break // nothing to hunt != everything demoted
  round++
+  const spentBefore = budget.spent()
+  const counts = { suppressedLedger: 0, suppressedRun: 0, finderNull: 0, finderEmpty: 0, verifierNull: 0 }
  const lenses = family.filter((l) => (cleanStreak[l.key] || 0) < 2)
  if (!lenses.length) {
    dry++
-    roundStats.push({ round, family: familyName(round), lenses: 0, candidates: 0, fresh: 0, confirmed: 0, refuted: 0, dryEligible: true, dryAfter: dry })
+    roundStats.push({ round, family: familyName(round), lenses: 0, candidates: 0, fresh: 0, confirmed: 0, refuted: 0, dryEligible: true, dryAfter: dry, severity: { critical: 0, high: 0, medium: 0, low: 0 }, perLens: {}, filesTouched: 0, filesNew: 0, ...counts, spentBefore, spentAfter: budget.spent() })
    log(`Round ${round}: every lens demoted - counts as a dry round (dry=${dry})`)
    continue
  }
@@ -476,123 +538,184 @@ while (dry < DRY_THRESHOLD && round < MAX_ROUNDS) {
  const lensResults = await pipeline(
    lenses,
    (lens) =>
-      parallel([
-        () => agent(finderPrompt(lens, rnd), { label: `r${rnd}:hunt:${lens.key}:opus`, phase: `Round ${rnd}`, model: 'opus', effort: 'high', schema: FINDINGS }),
-        () => agent(finderPrompt(lens, rnd), { label: `r${rnd}:hunt:${lens.key}:sonnet`, phase: `Round ${rnd}`, model: 'sonnet', effort: 'high', schema: FINDINGS }),
-      ]).then((pair) => ({ lens, pair })),
+      agent(finderPrompt(lens, rnd), { label: `r${rnd}:hunt:${lens.key}:opus`, phase: `Round ${rnd}`, model: 'opus', effort: 'high', schema: FINDINGS })
+        .then((res) => ({ lens, res })),
    async (r) => {
-      const { lens, pair } = r
-      const finderFailed = pair.some((p) => p === null)
-      // Tag by panel slot, not by position in the surviving list: filtering the nulls out first
-      // would shift sonnet into slot 0 whenever opus dies and mislabel its finds. The panel unions
-      // rather than votes, so this is the only signal for whether the second finder earns its cost
-      // - and since dedupe keeps the first occurrence and opus is slot 0, `finder: 'sonnet'` means
-      // opus did not report it. `finder: 'opus'` says nothing about sonnet either way.
-      const union = pair.flatMap((p, i) =>
-        p ? (p.findings || []).map((f) => ({ ...f, finder: i ? 'sonnet' : 'opus' })) : [],
-      )
-      const fresh = dedupe(union, seenAtStart)
-      if (!fresh.length) return { lens, finderFailed, unionCount: union.length, fresh: [], verdicts: [] }
+      const { lens, res } = r
+      // agent() returns null on failure; a thrown stage instead nulls the whole lens result,
+      // which the eligibility check catches separately. Both checks are needed.
+      const finderFailed = res === null
+      if (finderFailed) counts.finderNull++
+      else if (!(res.findings || []).length) counts.finderEmpty++
+      // finder is constant now; kept on the record for ledger continuity across hunts
+      const union = res ? (res.findings || []).map((f) => ({ ...f, finder: 'opus' })) : []
+      const fresh = dedupe(union, seenAtStart, counts)
+      if (!fresh.length) return { lens, finderFailed, unionCount: union.length, fresh: [], matched: [], unmatched: [] }
      log(`r${rnd} ${lens.key}: ${fresh.length} fresh candidate(s) -> verification`)
      const vopts = { phase: `Round ${rnd}`, model: 'fable', effort: 'high', schema: VERDICTS }
-      let v = await agent(verifyPrompt(lens.key, fresh), { ...vopts, label: `r${rnd}:verify:${lens.key}` })
-      if (!v || (v.verdicts || []).length < fresh.length) {
-        const retry = await agent(verifyPrompt(lens.key, fresh), { ...vopts, label: `r${rnd}:verify:${lens.key}:retry` })
-        if (((retry && retry.verdicts) || []).length > ((v && v.verdicts) || []).length) v = retry
+      // Pair verdicts to candidates as they arrive, then retry ONLY what got no usable verdict.
+      // Retrying the whole batch re-burned every verdict on a partial return, and the old
+      // count-based trigger let N unmatched garbage verdicts skip the retry entirely.
+      const matched = []
+      const unmatched = fresh.slice()
+      const absorb = (vs) => {
+        for (const v of vs || []) {
+          const vRec = { file: v.file, line: v.line, title: v.title }
+          const idx = unmatched.findIndex((f) => isDup(vRec, f) || isDup(f, vRec))
+          if (idx === -1) {
+            const claimed = matched.some(({ cand }) => isDup(vRec, cand) || isDup(cand, vRec))
+            log(`r${rnd} ${lens.key}: verifier verdict "${v.title}" (${v.file}:${v.line}) ${claimed ? 'duplicates an already-claimed candidate' : 'matched no candidate'} - dropped`)
+            continue
+          }
+          const [cand] = unmatched.splice(idx, 1)
+          matched.push({ v, cand })
+        }
      }
-      return { lens, finderFailed, unionCount: union.length, fresh, verdicts: (v && v.verdicts) || [] }
+      const v1 = await agent(verifyPrompt(lens.key, fresh), { ...vopts, label: `r${rnd}:verify:${lens.key}` })
+      absorb(v1 && v1.verdicts)
+      if (!v1) counts.verifierNull++
+      if (unmatched.length) {
+        const v2 = await agent(verifyPrompt(lens.key, unmatched.slice()), { ...vopts, label: `r${rnd}:verify:${lens.key}:retry` })
+        absorb(v2 && v2.verdicts)
+        if (!v2) counts.verifierNull++
+      }
+      return { lens, finderFailed, unionCount: union.length, fresh, matched, unmatched }
    },
  )

+  // a thrown stage nulls the whole lens result - rewind its explore draw too, or the
+  // session records never-read files as explored-clean (the same poison as a null finder)
+  lensResults.forEach((r, i) => {
+    if (!r && lenses[i].files) for (const f of lenses[i].files) exploreConsumed.delete(f)
+  })
+
  let eligible = !lensResults.some((r) => !r)
  let newConfirmed = 0
  let newRefuted = 0
  let candCount = 0
  let freshCount = 0
+  const perLens = {}
+  const sevMix = { critical: 0, high: 0, medium: 0, low: 0 }
+  const filesTouched = new Set()
+  const filesNew = new Set()
  for (const r of lensResults.filter(Boolean)) {
    candCount += r.unionCount
    freshCount += r.fresh.length
    if (r.finderFailed) eligible = false
+    if (r.finderFailed && r.lens.files) {
+      // a dead finder read nothing: un-consume its draw so later rounds can re-offer the
+      // files and the session does not record never-examined files as explored-clean
+      for (const f of r.lens.files) exploreConsumed.delete(f)
+    }
    let lensConfirmed = 0
-    const unmatched = r.fresh.slice()
-    for (const v of r.verdicts) {
-      const vRec = { file: v.file, line: v.line, title: v.title }
-      const idx = unmatched.findIndex((f) => isDup(vRec, f) || isDup(f, vRec))
-      if (idx === -1) {
-        log(`r${round} ${r.lens.key}: verifier verdict "${v.title}" (${v.file}:${v.line}) matched no candidate - dropped`)
-        continue
-      }
+    let lensRefuted = 0
+    for (const { v, cand } of r.matched) {
      // Keep the matched candidate: the verdict schema has no why/repro/evidence, and the
      // ledger needs them. Verdict fields are spread last so the verifier's re-rated severity
      // and its corrected title/file/line win over the finder's.
-      const [cand] = unmatched.splice(idx, 1)
      const rec = { file: v.file, line: v.line, title: v.title, status: v.refuted ? 'refuted' : 'confirmed' }
-      if (seen.some((p) => isDup(rec, p))) continue // cross-lens same-round duplicate
+      if (seen.some((p) => isDup(rec, p))) { counts.suppressedRun++; continue } // cross-lens same-round duplicate
      seen.push(rec)
-      if (v.refuted) newRefuted++
+      if (v.refuted) { newRefuted++; lensRefuted++ }
      else {
        newConfirmed++
        lensConfirmed++
+        sevMix[v.severity] = (sevMix[v.severity] || 0) + 1
        confirmedAll.push({ ...cand, ...v, lens: r.lens.key, round })
      }
    }
-    if (unmatched.length) {
+    if (r.unmatched.length) {
      eligible = false // partial verifier failure: some candidates got no verdict at all
-      for (const f of unmatched) unverified.push({ ...f, lens: r.lens.key, round })
+      for (const f of r.unmatched) unverified.push({ ...f, lens: r.lens.key, round })
    }
    // a lens hunted at partial panel strength, or whose candidates never got a verdict, is not evidence of cleanliness
-    if (!r.finderFailed && !unmatched.length) cleanStreak[r.lens.key] = lensConfirmed > 0 ? 0 : (cleanStreak[r.lens.key] || 0) + 1
+    if (!r.finderFailed && !r.unmatched.length) cleanStreak[r.lens.key] = lensConfirmed > 0 ? 0 : (cleanStreak[r.lens.key] || 0) + 1
+    perLens[r.lens.key] = { candidates: r.unionCount, fresh: r.fresh.length, confirmed: lensConfirmed, refuted: lensRefuted, unverified: r.unmatched.length }
+    // coverage proxy: files that produced fresh candidates this round (finder reading is unobservable)
+    for (const f of r.fresh) {
+      filesTouched.add(f.file)
+      if (!seenAtStart.some((s) => s.file === f.file)) filesNew.add(f.file)
+    }
  }

  if (newConfirmed > 0) dry = 0
  else if (eligible) dry++
  // ineligible zero-confirm round: dry unchanged - "we didn't fully look" is not "it's clean"
-  roundStats.push({ round, family: familyName(round), lenses: lenses.length, candidates: candCount, fresh: freshCount, confirmed: newConfirmed, refuted: newRefuted, dryEligible: eligible, dryAfter: dry })
+  roundStats.push({ round, family: familyName(round), lenses: lenses.length, candidates: candCount, fresh: freshCount, confirmed: newConfirmed, refuted: newRefuted, dryEligible: eligible, dryAfter: dry, severity: sevMix, perLens, filesTouched: filesTouched.size, filesNew: filesNew.size, ...counts, spentBefore, spentAfter: budget.spent() })
  log(`Round ${round} (${familyName(round)}): ${newConfirmed} confirmed, ${newRefuted} refuted, dry=${dry}${eligible ? '' : ' (ineligible)'}`)
 }

 const converged = dry >= DRY_THRESHOLD

-// The panel unions rather than votes, so the second finder's whole value is what it finds alone.
-// dedupe keeps the opus-slot record when both report the same bug, so a confirmed finding tagged
-// sonnet is one opus missed. A run where that count is 0 is the evidence for dropping the second
-// model; anything above 0 is what it bought.
-const sonnetOnly = confirmedAll.filter((f) => f.finder === 'sonnet').length
-log(`panel: ${confirmedAll.length} confirmed, ${sonnetOnly} sonnet-only (opus missed), ${confirmedAll.length - sonnetOnly} found by opus`)
-
-// ---------- report ----------
-phase('Report')
+// ---------- report (deterministic) ----------
+// A report agent silently dropped findings (79 sections for 82 confirmed on 2026-08-12), so the
+// markdown is assembled in-script from confirmedSorted. Coordinate spot-checking moved to the
+// calling session, which validates EVERY finding's file/line after return (see bughunt-run).
 const RANK = { critical: 0, high: 1, medium: 2, low: 3 }
 const confirmedSorted = confirmedAll.slice().sort((a, b) => RANK[a.severity] - RANK[b.severity])
 const unverifiedFinal = unverified.filter((u) => !seen.some((p) => isDup(u, p)))
 const table = convergenceTable(roundStats, converged, stoppedOnBudget)
-
-let report
-if (!confirmedSorted.length && !unverifiedFinal.length) {
-  const outcome = converged ? 'Converged' : stoppedOnBudget ? 'Stopped on budget - NOT converged' : 'Hit the round backstop - NOT converged'
-  report = `${outcome} after ${round} round(s) with zero confirmed findings.\n\n${table}`
-} else {
-  report = await agent(
-    `You are writing the final bug-hunt report for OwnCord (D:/Local-Lab/Repos/OwnCord).\n\n` +
-      `The findings below already survived adversarial verification - do NOT re-litigate them, and do NOT add new ` +
-      `ones. Your job is presentation and prioritization for a maintainer who will fix these today.\n\n` +
-      `Spot-check the two highest-severity findings against the real files to make sure file paths and line numbers ` +
-      `are accurate; correct them silently if they drifted.\n\n` +
-      `Write markdown:\n` +
-      `  - Open with one paragraph: how many real bugs, in which subsystems, whether the hunt CONVERGED, and which ` +
-      `finding to fix first and why.\n` +
-      `  - Then one section per finding, ordered by severity: a "### <severity> - <title>" heading, the ` +
-      `\`file:line\` reference, what breaks and under exactly what conditions, and the smallest correct fix.\n` +
-      (unverifiedFinal.length
-        ? `  - Then an "## Unverified - re-run" section listing these candidates whose verification failed twice: ` +
-          `${JSON.stringify(unverifiedFinal)}\n`
-        : '') +
-      `  - End with the convergence table below, VERBATIM.\n` +
-      `Write in complete sentences. No emoji, no "consider" hedging.\n\n` +
-      `--- CONFIRMED FINDINGS ---\n${JSON.stringify(confirmedSorted, null, 2)}\n\n` +
-      `--- CONVERGENCE TABLE ---\n${table}`,
-    { label: 'report', phase: 'Report', model: 'fable', effort: 'high' },
-  )
+const sum = (k) => roundStats.reduce((n, r) => n + (r[k] || 0), 0)
+const runStats = {
+  config: { maxRounds: MAX_ROUNDS, dryThreshold: DRY_THRESHOLD, customLenses: !!CUSTOM_LENSES, knownCount: (ARGS.known || []).length, graphRows: GRAPH_ROWS.length, budgetTotal: budget.total },
+  spentTotal: budget.spent(),
+  rounds: roundStats.length,
+  converged,
+  stoppedOnBudget,
+  confirmed: confirmedSorted.length,
+  refuted: sum('refuted'),
+  unverified: unverifiedFinal.length,
+  suppressedLedger: sum('suppressedLedger'),
+  suppressedRun: sum('suppressedRun'),
+  finderNull: sum('finderNull'),
+  finderEmpty: sum('finderEmpty'),
+  verifierNull: sum('verifierNull'),
 }

-return { converged, stoppedOnBudget, rounds: roundStats, confirmed: confirmedSorted, unverified: unverifiedFinal, report }
+function buildReport() {
+  const outcome = converged
+    ? `CONVERGED after ${round} round(s).`
+    : stoppedOnBudget
+      ? `NOT converged - stopped on budget after ${round} round(s).`
+      : `NOT converged - hit the round backstop after ${round} round(s).`
+  const sev = { critical: 0, high: 0, medium: 0, low: 0 }
+  for (const f of confirmedSorted) sev[f.severity] = (sev[f.severity] || 0) + 1
+  const lines = ['# Bug hunt report', '']
+  lines.push(
+    `${confirmedSorted.length} confirmed finding(s) - ${sev.critical} critical, ${sev.high} high, ` +
+      `${sev.medium} medium, ${sev.low} low. ${outcome}` +
+      (confirmedSorted.length
+        ? ` Fix first: ${confirmedSorted[0].title} (\`${confirmedSorted[0].file}:${confirmedSorted[0].line}\`).`
+        : ''),
+    '',
+  )
+  for (const f of confirmedSorted) {
+    lines.push(`### ${f.severity} - ${f.title}`, '')
+    lines.push(`\`${f.file}:${f.line}\` - lens \`${f.lens}\`, round ${f.round}, confidence ${f.confidence}`, '')
+    if (f.why) lines.push(f.why, '')
+    if (f.repro) lines.push(`**Repro:** ${f.repro}`, '')
+    if (f.evidence) lines.push(`**Evidence:** ${f.evidence}`, '')
+    if (f.fix) lines.push(`**Fix:** ${f.fix}`, '')
+  }
+  if (unverifiedFinal.length) {
+    lines.push('## Unverified - re-run', '')
+    for (const u of unverifiedFinal) lines.push(`- \`${u.file}:${u.line}\` ${u.title} (lens \`${u.lens}\`, round ${u.round})`)
+    lines.push('')
+  }
+  lines.push(table)
+  lines.push('', '## Run stats', '')
+  lines.push(
+    `Total spent: ${runStats.spentTotal} output tokens across ${runStats.rounds} round(s). ` +
+      `Suppressed by dedupe: ${runStats.suppressedLedger} ledger-known, ${runStats.suppressedRun} same-run. ` +
+      `Agent failures: ${runStats.finderNull} finder null, ${runStats.finderEmpty} finder empty, ${runStats.verifierNull} verifier null.`,
+    '',
+  )
+  lines.push('| round | spent | files (new) | suppressed ledger/run | finder null/empty | verifier null |')
+  lines.push('|---|---|---|---|---|---|')
+  for (const s of roundStats)
+    lines.push(`| ${s.round} | ${s.spentAfter - s.spentBefore} | ${s.filesTouched} (${s.filesNew}) | ${s.suppressedLedger}/${s.suppressedRun} | ${s.finderNull}/${s.finderEmpty} | ${s.verifierNull} |`)
+  return lines.join('\n')
+}
+const report = buildReport()
+
+return { converged, stoppedOnBudget, rounds: roundStats, confirmed: confirmedSorted, unverified: unverifiedFinal, runStats, exploredFiles: [...exploreConsumed], report }
@@ -98,3 +98,4 @@ Client/tauri-client/.env

 # local server run logs
 server.log
+graphify-out/