@oneie/claude 0.7.0 → 0.8.0

package/scripts/do-derive-check.sh

@@ -0,0 +1,436 @@
+#!/usr/bin/env bash
+# do-derive-check.sh — the five checks that hold the agent-native promise.
+#
+# Contract: text/agent-native-plan.md § 3.5. Six of the promise's eight proof
+# legs are this script, so that section is the specification and this file is
+# the implementation of it.
+#
+#   contract   60+ receivers advertised to MCP
+#   mcp        the derived tool list is computed, complete, and schema-faithful
+#   binder     an unauthorised caller is stopped BY THE WRAPPER, affordance kept
+#   routes     no new JSON endpoint duplicates a receiver reachable by name
+#   counts     no hand-written tool ships beside its derived twin
+#
+# THREE EXIT CODES, NOT TWO:
+#   0  ok          delivered and behaving
+#   1  RED         genuinely missing or wrong
+#   3  CANNOT RUN  the check could not reach its evidence
+#
+# Red and cannot-run send a person to opposite places. `contract` and `mcp`
+# import TypeScript through bun; an import that fails because a package is
+# unbuilt is CANNOT RUN, not deliverable-missing — the difference between "go
+# build the SDK" and "go re-scope the work".
+#
+# Every threshold is env-overridable. That is not a convenience: --self-test
+# drives each check to red by moving its threshold against the real tree, which
+# proves the comparison is wired rather than that a fixture can be written.
+#
+# monorepo-only
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+
+MIN_MCP_FLAGGED="${DERIVE_MIN_MCP_FLAGGED:-60}"
+MIN_DERIVED_TOOLS="${DERIVE_MIN_DERIVED_TOOLS:-60}"
+MAX_HANDWRITTEN_MCP="${DERIVE_MAX_HANDWRITTEN_MCP:-95}"   # strictly below
+MAX_CURATED_CHAT="${DERIVE_MAX_CURATED_CHAT:-125}"        # at or below
+API_DIR="${DERIVE_API_DIR:-$ROOT/one.ie/web/src/pages/api}"
+MCP_TOOLS_DIR="${DERIVE_MCP_TOOLS_DIR:-$ROOT/packages/mcp/src/tools}"
+CHAT_TOOLS_DIR="${DERIVE_CHAT_TOOLS_DIR:-$ROOT/channels/src/tools}"
+ALLOWLIST="${DERIVE_ROUTE_ALLOWLIST:-$ROOT/.claude/derive-route-allowlist.txt}"
+
+say() { printf '[derive-check] %s\n' "$*"; }
+red() { printf '[derive-check] RED — %s\n' "$*" >&2; return 1; }
+cant() { printf '[derive-check] CANNOT RUN — %s\n' "$*" >&2; return 3; }
+
+need_bun() { command -v bun >/dev/null || return 3; return 0; }
+
+# ── contract ─────────────────────────────────────────────────────────────────
+# The Receiver contract carries surfaces.mcp, populated.
+check_contract() {
+  need_bun || { cant "bun is not on PATH"; return 3; }
+  local src="$ROOT/packages/sdk/src/receivers.ts"
+  [ -f "$src" ] || { cant "receivers.ts absent at $src"; return 3; }
+
+  local out rc
+  out="$(cd "$ROOT" && bun -e "
+    import('./packages/sdk/src/receivers.ts')
+      .then(m => {
+        const R = Object.values(m.RECEIVERS);
+        const flagged = R.filter(r => r.surfaces && r.surfaces.mcp);
+        const named = flagged.filter(r => typeof r.surfaces.mcp === 'object' && r.surfaces.mcp.name);
+        const dupes = named.length - new Set(named.map(r => r.surfaces.mcp.name)).size;
+        console.log(JSON.stringify({ total: R.length, flagged: flagged.length, named: named.length, dupes }));
+      })
+      .catch(e => { console.error('IMPORT ' + e); process.exit(3); });
+  " 2>&1)"; rc=$?
+
+  [ "$rc" -eq 3 ] && { cant "receivers.ts did not import: ${out:0:160}"; return 3; }
+  [ "$rc" -ne 0 ] && { cant "bun exited $rc: ${out:0:160}"; return 3; }
+
+  local flagged named dupes total
+  flagged=$(printf '%s' "$out" | grep -o '"flagged":[0-9]*' | cut -d: -f2)
+  named=$(printf '%s' "$out" | grep -o '"named":[0-9]*' | cut -d: -f2)
+  dupes=$(printf '%s' "$out" | grep -o '"dupes":[0-9]*' | cut -d: -f2)
+  total=$(printf '%s' "$out" | grep -o '"total":[0-9]*' | cut -d: -f2)
+  [ -z "$flagged" ] && { cant "could not parse the receiver census: ${out:0:160}"; return 3; }
+
+  # A duplicate declared name would silently shadow one tool with another.
+  [ "${dupes:-0}" -gt 0 ] && { red "$dupes duplicate declared MCP tool name(s)"; return 1; }
+  [ "$flagged" -lt "$MIN_MCP_FLAGGED" ] && {
+    red "surfaces.mcp on $flagged receivers, need >= $MIN_MCP_FLAGGED"; return 1; }
+
+  say "contract ok — $flagged/$total flagged ($named carry an existing tool name)"
+  return 0
+}
+
+# ── mcp ──────────────────────────────────────────────────────────────────────
+# The tool list is COMPUTED from the registry, complete, and schema-faithful.
+check_mcp() {
+  need_bun || { cant "bun is not on PATH"; return 3; }
+  local src="$ROOT/packages/mcp/src/tools/from-registry.ts"
+  local dist="$ROOT/packages/mcp/dist/index.js"
+  [ -f "$src" ] || { red "from-registry.ts absent — the derivation was never written"; return 1; }
+  [ -f "$dist" ] || { cant "packages/mcp is unbuilt (no dist/index.js) — run bun run build"; return 3; }
+
+  local out rc
+  out="$(cd "$ROOT/packages/mcp" && bun -e "
+    import('./dist/index.js')
+      .then(async m => {
+        const tools = m.mcpToolsFromRegistry();
+        const meta = await import('@oneie/sdk/meta');
+        const { RECEIVERS } = await import('@oneie/sdk/receivers');
+        // Schema fidelity: a derived tool must advertise ITS OWN receiver's request
+        // schema. A tool carrying the wrong schema passes every count and lies to
+        // every caller, so this is checked per tool, not sampled.
+        let wrong = 0;
+        for (const r of Object.values(RECEIVERS)) {
+          if (!r.surfaces || !r.surfaces.mcp) continue;
+          const want = m.toolNameFor(r.receiver, r.surfaces.mcp);
+          const t = tools.find(x => x.name === want);
+          if (!t) { wrong++; continue; }
+          const expect = JSON.stringify(meta.metaSchema(r.receiver).request);
+          if (JSON.stringify(t.inputSchema) !== expect) wrong++;
+        }
+        const noSchema = tools.filter(t => !t.inputSchema || typeof t.inputSchema !== 'object').length;
+        console.log(JSON.stringify({ derived: tools.length, unconvertible: m.UNCONVERTIBLE.length, wrong, noSchema }));
+      })
+      .catch(e => { console.error('IMPORT ' + e); process.exit(3); });
+  " 2>&1)"; rc=$?
+
+  [ "$rc" -eq 3 ] && { cant "mcp dist did not import: ${out:0:160}"; return 3; }
+  [ "$rc" -ne 0 ] && { cant "bun exited $rc: ${out:0:160}"; return 3; }
+
+  local derived unconv wrong noschema
+  derived=$(printf '%s' "$out" | grep -o '"derived":[0-9]*' | cut -d: -f2)
+  unconv=$(printf '%s' "$out" | grep -o '"unconvertible":[0-9]*' | cut -d: -f2)
+  wrong=$(printf '%s' "$out" | grep -o '"wrong":[0-9]*' | cut -d: -f2)
+  noschema=$(printf '%s' "$out" | grep -o '"noSchema":[0-9]*' | cut -d: -f2)
+  [ -z "$derived" ] && { cant "could not parse the derivation census: ${out:0:160}"; return 3; }
+
+  [ "${unconv:-0}" -gt 0 ] && { red "$unconv receiver schema(s) could not convert to JSON Schema"; return 1; }
+  [ "${noschema:-0}" -gt 0 ] && { red "$noschema derived tool(s) carry no inputSchema"; return 1; }
+  [ "${wrong:-0}" -gt 0 ] && { red "$wrong derived tool(s) missing or advertising the wrong receiver's schema"; return 1; }
+  [ "$derived" -lt "$MIN_DERIVED_TOOLS" ] && {
+    red "$derived derived tools, need >= $MIN_DERIVED_TOOLS"; return 1; }
+
+  say "mcp ok — $derived derived tools, every schema matches its receiver"
+  return 0
+}
+
+# ── binder ───────────────────────────────────────────────────────────────────
+# Behavioural first, structural second. Counting handler tables is true the
+# moment everything routes through one function whether or not that function
+# ever consults the policy — so the count is the WEAKER half and runs last.
+check_binder() {
+  need_bun || { cant "bun is not on PATH"; return 3; }
+  local f="$ROOT/one.ie/web/src/lib/bind-receiver.ts"
+  [ -f "$f" ] || { red "bind-receiver.ts absent — the wrapper was never written"; return 1; }
+
+  local out rc
+  out="$(cd "$ROOT/one.ie/web" && bun -e "
+    import('./src/lib/bind-receiver.ts')
+      .then(async m => {
+        const fail = [];
+
+        // A1 — a tenant-label receiver, called by a caller holding nothing, is
+        //      rejected BY THE WRAPPER and the handler never runs.
+        // Assert the OUTCOME (refused, handler never reached), not the mechanism.
+        // The floor returns a refusal object rather than throwing, because that
+        // is how every handler in the codebase signals refusal and what the ask
+        // route already reads. A check that demanded a throw would fail a
+        // correct implementation.
+        let ran = false;
+        const bound = m.bindReceiver('world:create-group', async () => { ran = true; return { ok: true }; });
+        let refused = false;
+        try {
+          const r = await bound({ name: 'x', slug: 'y' }, {}, {});
+          refused = !!(r && typeof r === 'object' && (r.forbidden === true || r.ok === false));
+        } catch { refused = true; }
+        if (!refused) fail.push('A1: unauthorised caller was NOT refused');
+        if (ran) fail.push('A1: handler RAN despite the auth gate — the wrapper consulted the policy and ignored it');
+
+        // A2 — the error-as-affordance body still exists and still teaches.
+        //      Checked at its OWNER, receiver-envelope, not at the binder: the
+        //      binder's request check is warn-only (the declared schemas had
+        //      drifted from real in-process callers), while /api/ask and
+        //      /api/signal still reject with this exact body. Asserting it on
+        //      the binder would assert a contract this cycle deliberately does
+        //      not ship; asserting it here checks the one that does.
+        const env = await import('./src/lib/receiver-envelope.ts');
+        const v = env.validateReceiver('world:create-group', {});
+        if (v.ok) {
+          fail.push('A2: an invalid request validated clean — the affordance gate is gone');
+        } else {
+          const body = v.body || {};
+          if (!('hint' in body) || !('expected' in body) || !('error' in body)) {
+            fail.push('A2: affordance body lost a teaching field: ' + JSON.stringify(body).slice(0, 120));
+          }
+        }
+        // A2b — and the binder must still NOTICE the mismatch, or warn-only has
+        //       silently become no-op.
+        const seenReq = [];
+        const rw = console.warn;
+        console.warn = (...a) => seenReq.push(a.join(' '));
+        const b2 = m.bindReceiver('world:create-group', async () => ({ ok: true }));
+        await b2({}, {}, { ownerSlug: 'someone' }).catch(() => {});
+        console.warn = rw;
+        if (!seenReq.some(s => /request-mismatch/.test(s))) {
+          fail.push('A2b: an invalid request produced no request-mismatch warning — warn-only became no-op');
+        }
+
+        // A3 — a declared response is actually consulted. WARN mode counts, but
+        //      the mismatch must be OBSERVABLE, not swallowed.
+        const seen = [];
+        const realWarn = console.warn;
+        console.warn = (...a) => seen.push(a.join(' '));
+        const b3 = m.bindReceiver('world:create-group', async () => ({ nonsense: true }));
+        await b3({ name: 'x', slug: 'y' }, {}, { ownerSlug: 'someone', staff: true }).catch(() => {});
+        console.warn = realWarn;
+        if (!seen.some(s => /response-mismatch/.test(s))) {
+          fail.push('A3: a handler returning the wrong shape produced no response-mismatch warning');
+        }
+
+        // A4 — an unlisted auth label must never fall OPEN.
+        if (m.authClassFor('a-label-that-does-not-exist') === 'open') {
+          fail.push('A4: an unknown auth label resolves to open');
+        }
+
+        console.log(JSON.stringify({ fail }));
+      })
+      .catch(e => { console.error('IMPORT ' + e); process.exit(3); });
+  " 2>&1)"; rc=$?
+
+  [ "$rc" -eq 3 ] && { cant "bind-receiver.ts did not import: ${out:0:200}"; return 3; }
+  [ "$rc" -ne 0 ] && { cant "bun exited $rc: ${out:0:200}"; return 3; }
+
+  local fails
+  fails="$(printf '%s' "$out" | sed -n 's/.*"fail":\[\(.*\)\]}.*/\1/p')"
+  [ -n "$fails" ] && [ "$fails" != "" ] && { red "binder behaviour: $fails"; return 1; }
+
+  # Structural, and deliberately last: a raw handler table that never mentions
+  # bindReceiver is a dispatcher that has to remember.
+  local raw=0 f2
+  while IFS= read -r f2; do
+    grep -q "bindReceiver" "$f2" || { say "  raw handler table: ${f2#$ROOT/}"; raw=$((raw + 1)); }
+  done < <(grep -rlE '^const (HANDLERS|RESOLVERS)' "$ROOT/one.ie/web/src/lib" --include='*.ts' 2>/dev/null || true)
+  [ "$raw" -gt 0 ] && { red "$raw handler table(s) bind outside bindReceiver"; return 1; }
+
+  say "binder ok — rejects before the handler, keeps the affordance, checks the response, 0 raw tables"
+  return 0
+}
+
+# ── routes ───────────────────────────────────────────────────────────────────
+# A new JSON endpoint duplicating a receiver reachable by name. NOT a cap on
+# routes: auth callbacks, webhooks, uploads and non-JSON responses are exempt,
+# and today's tree is grandfathered in an allowlist that may shrink, never grow.
+EXEMPT_RE='(^|/)(auth|webhook|webhooks|upload|uploads|oauth|callback|stripe|sse|stream|rss|sitemap|feed|og|image|images|avatar|media|export|cron|health)(/|$)'
+
+check_routes() {
+  [ -d "$API_DIR" ] || { cant "api dir absent at $API_DIR"; return 3; }
+  [ -f "$ALLOWLIST" ] || { cant "route allowlist absent at $ALLOWLIST — freeze it first (--freeze-routes)"; return 3; }
+  need_bun || { cant "bun is not on PATH"; return 3; }
+
+  local receivers
+  receivers="$(cd "$ROOT" && bun -e "
+    import('./packages/sdk/src/receivers.ts')
+      .then(m => console.log(Object.keys(m.RECEIVERS).join('\n')))
+      .catch(() => process.exit(3));
+  " 2>/dev/null)" || { cant "could not read the receiver list"; return 3; }
+  [ -z "$receivers" ] && { cant "receiver list came back empty"; return 3; }
+
+  # noun set: the second half of every receiver name, singularised
+  local nouns
+  nouns="$(printf '%s\n' "$receivers" | sed 's/^[a-z-]*://' | sed 's/s$//' | sort -u)"
+
+  local offenders=0 rel key
+  while IFS= read -r f; do
+    rel="${f#$API_DIR/}"
+    printf '%s\n' "$rel" | grep -qE "$EXEMPT_RE" && continue
+    grep -qF -- "$rel" "$ALLOWLIST" && continue
+    # op key: path minus dynamic segments and extension, singularised
+    key="$(printf '%s' "$rel" | sed 's/\.ts$//' | sed 's/\[[^]]*\]//g' | tr '/' ' ' | tr -s ' ')"
+    for n in $key; do
+      n="$(printf '%s' "$n" | sed 's/s$//')"
+      [ -z "$n" ] && continue
+      if printf '%s\n' "$nouns" | grep -qx -- "$n"; then
+        say "  twin: $rel duplicates a receiver reachable by name (noun '$n')"
+        offenders=$((offenders + 1))
+        break
+      fi
+    done
+  done < <(find "$API_DIR" -name '*.ts' -type f 2>/dev/null)
+
+  [ "$offenders" -gt 0 ] && {
+    red "$offenders new route(s) duplicate a by-name receiver — use the door, or add an exempt kind"; return 1; }
+
+  say "routes ok — no new twin against $(wc -l < "$ALLOWLIST" | tr -d ' ') grandfathered entries"
+  return 0
+}
+
+freeze_routes() {
+  [ -d "$API_DIR" ] || { cant "api dir absent at $API_DIR"; return 3; }
+  find "$API_DIR" -name '*.ts' -type f | sed "s|^$API_DIR/||" | sort > "$ALLOWLIST"
+  say "froze $(wc -l < "$ALLOWLIST" | tr -d ' ') route(s) into ${ALLOWLIST#$ROOT/} — this list may shrink, never grow"
+}
+
+# ── counts ───────────────────────────────────────────────────────────────────
+# The deletion, measured. The load-bearing assertion is COEXISTENCE, not the
+# number: "strictly below 95" alone is satisfied by deleting one tool.
+check_counts() {
+  [ -d "$MCP_TOOLS_DIR" ] || { cant "mcp tools dir absent at $MCP_TOOLS_DIR"; return 3; }
+  need_bun || { cant "bun is not on PATH"; return 3; }
+
+  local flagged
+  flagged="$(cd "$ROOT" && bun -e "
+    import('./packages/sdk/src/receivers.ts')
+      .then(m => {
+        for (const r of Object.values(m.RECEIVERS)) if (r.surfaces && r.surfaces.mcp) console.log(r.receiver);
+      })
+      .catch(() => process.exit(3));
+  " 2>/dev/null)" || { cant "could not read the flagged receivers"; return 3; }
+
+  # Coexistence: a flagged receiver must not still have a SINGLE-RECEIVER
+  # hand-written wrapper. Scoped to single-receiver tools on purpose — a
+  # surviving multi-receiver recipe legitimately names a flagged receiver among
+  # several, and the pragma covers anything this still catches wrongly.
+  local coexist=0 r hits
+  while IFS= read -r r; do
+    [ -z "$r" ] && continue
+    while IFS= read -r tf; do
+      grep -qF -- 'derive:allow-handwritten' "$tf" && continue
+      # a module naming exactly ONE distinct receiver is a single-receiver wrapper
+      hits="$(grep -ohE '"[a-z][a-z-]*:[a-z][a-z-]*"' "$tf" 2>/dev/null | sort -u | wc -l | tr -d ' ')"
+      [ "${hits:-0}" -eq 1 ] || continue
+      say "  coexists: $r still has a hand-written wrapper in ${tf#$ROOT/}"
+      coexist=$((coexist + 1))
+    done < <(grep -rlF -- "\"$r\"" "$MCP_TOOLS_DIR" --include='*.ts' 2>/dev/null || true)
+  done <<< "$flagged"
+
+  [ "$coexist" -gt 0 ] && { red "$coexist hand-written tool(s) ship beside a derived twin"; return 1; }
+
+  # from-registry.ts is the DERIVATION, not a hand-written tool. Counting it
+  # inflated the census by one and would have let the file that deletes
+  # hand-written tools count as one itself.
+  local hw cur
+  hw="$(grep -h 'name: "' $(ls "$MCP_TOOLS_DIR"/*.ts 2>/dev/null | grep -v '/from-registry\.ts$') 2>/dev/null | wc -l | tr -d ' ')"
+  [ "$hw" -ge "$MAX_HANDWRITTEN_MCP" ] && {
+    red "$hw hand-written MCP tool definitions, must be strictly below $MAX_HANDWRITTEN_MCP"; return 1; }
+
+  if [ -d "$CHAT_TOOLS_DIR" ]; then
+    cur="$(grep -rho 'defineWorkspaceTool(' "$CHAT_TOOLS_DIR"/*.ts 2>/dev/null | wc -l | tr -d ' ')"
+    [ "${cur:-0}" -gt "$MAX_CURATED_CHAT" ] && {
+      red "$cur curated chat tools, ceiling is $MAX_CURATED_CHAT"; return 1; }
+  else
+    cur="n/a"
+  fi
+
+  say "counts ok — no coexistence, $hw hand-written MCP defs (< $MAX_HANDWRITTEN_MCP), $cur curated chat tools"
+  return 0
+}
+
+# ── self-test ────────────────────────────────────────────────────────────────
+# Each check is driven to RED against the real tree by moving its own threshold
+# or pointing it at a fixture carrying the defect. A checker that has only ever
+# run against a healthy tree proves nothing about what it would catch.
+self_test() {
+  local fails=0 rc tmp
+  tmp="$(mktemp -d)"; trap 'rm -rf "$tmp"' RETURN
+
+  expect_red() { # expect_red <label> <cmd...>
+    local label="$1"; shift
+    "$@" >/dev/null 2>&1; rc=$?
+    if [ "$rc" -eq 1 ]; then printf '  ok   %s -> RED\n' "$label"
+    else printf '  FAIL %s -> expected 1, got %s\n' "$label" "$rc"; fails=$((fails + 1)); fi
+  }
+  expect_cant() { # expect_cant <label> <cmd...>
+    local label="$1"; shift
+    "$@" >/dev/null 2>&1; rc=$?
+    if [ "$rc" -eq 3 ]; then printf '  ok   %s -> CANNOT RUN\n' "$label"
+    else printf '  FAIL %s -> expected 3, got %s\n' "$label" "$rc"; fails=$((fails + 1)); fi
+  }
+
+  # contract: an unreachable floor must go red, not pass
+  expect_red "contract (floor 99999)" env DERIVE_MIN_MCP_FLAGGED=99999 bash "$0" contract
+  # mcp: same
+  expect_red "mcp (floor 99999)" env DERIVE_MIN_DERIVED_TOOLS=99999 bash "$0" mcp
+  # counts: a ceiling of 0 hand-written defs must go red while any survive
+  expect_red "counts (max-handwritten 1)" env DERIVE_MAX_HANDWRITTEN_MCP=1 bash "$0" counts
+  # routes: a fixture tree carrying a twin of a real receiver
+  mkdir -p "$tmp/api/groups"
+  : > "$tmp/api/groups/create.ts"
+  : > "$tmp/allow.txt"
+  expect_red "routes (fixture twin)" env DERIVE_API_DIR="$tmp/api" DERIVE_ROUTE_ALLOWLIST="$tmp/allow.txt" bash "$0" routes
+  # routes: the same twin, allowlisted, must pass — the ratchet grandfathers
+  printf 'groups/create.ts\n' > "$tmp/allow.txt"
+  env DERIVE_API_DIR="$tmp/api" DERIVE_ROUTE_ALLOWLIST="$tmp/allow.txt" bash "$0" routes >/dev/null 2>&1
+  rc=$?
+  if [ "$rc" -eq 0 ]; then printf '  ok   routes (allowlisted) -> ok\n'
+  else printf '  FAIL routes (allowlisted) -> expected 0, got %s\n' "$rc"; fails=$((fails + 1)); fi
+  # cannot-run is distinguishable from red
+  expect_cant "routes (missing allowlist)" env DERIVE_API_DIR="$tmp/api" DERIVE_ROUTE_ALLOWLIST="$tmp/nope.txt" bash "$0" routes
+  expect_cant "counts (missing tools dir)" env DERIVE_MCP_TOOLS_DIR="$tmp/nope" bash "$0" counts
+
+  # binder: the defining fixture — a wrapper that consults the policy and calls
+  # the handler anyway passes every structural count and must fail A1.
+  cat > "$tmp/leaky.mjs" <<'EOF'
+const fail = [];
+let ran = false;
+const leaky = (receiver, handler) => async (d, e, c) => {
+  const authed = c && (c.staff === true || c.ownerSlug);   // consulted...
+  if (!authed) console.warn('would reject');               // ...and ignored
+  return handler(d, e, c);
+};
+const bound = leaky('world:create-group', async () => { ran = true; return { ok: true }; });
+let threw = false;
+try { await bound({}, {}, {}); } catch { threw = true; }
+if (!threw) fail.push('A1: unauthorised caller was NOT rejected');
+if (ran) fail.push('A1: handler RAN despite the auth gate');
+process.exit(fail.length ? 1 : 0);
+EOF
+  node "$tmp/leaky.mjs" >/dev/null 2>&1; rc=$?
+  if [ "$rc" -eq 1 ]; then printf '  ok   binder (leaky-wrapper fixture) -> RED\n'
+  else printf '  FAIL binder (leaky-wrapper fixture) -> expected 1, got %s\n' "$rc"; fails=$((fails + 1)); fi
+
+  if [ "$fails" -eq 0 ]; then say "self-test PASS — every check can go red, and red is distinguishable from cannot-run"; return 0; fi
+  printf '[derive-check] self-test FAILED (%s)\n' "$fails" >&2
+  return 1
+}
+
+# ── main ─────────────────────────────────────────────────────────────────────
+case "${1:-}" in
+  contract)       check_contract ;;
+  mcp)            check_mcp ;;
+  binder)         check_binder ;;
+  routes)         check_routes ;;
+  counts)         check_counts ;;
+  --freeze-routes) freeze_routes ;;
+  --self-test)    self_test ;;
+  all|"")
+    rc=0
+    for c in contract mcp binder routes counts; do
+      bash "$0" "$c" || rc=$?
+    done
+    exit "$rc" ;;
+  *) printf 'usage: %s {contract|mcp|binder|routes|counts|all|--self-test|--freeze-routes}\n' "$(basename "$0")" >&2; exit 2 ;;
+esac

package/scripts/do-fleet.sh

@@ -57,9 +57,33 @@ done
 # Cap slots by (cores - 2), never exceed --slots.
 CORES=$(getconf _NPROCESSORS_ONLN 2>/dev/null || echo 4)
 CORE_CAP=$(( CORES - 2 )); [ "$CORE_CAP" -lt 1 ] && CORE_CAP=1
-[ "$SLOTS" -gt "$CORE_CAP" ] && SLOTS="$CORE_CAP"
+BIND="--slots"
+[ "$SLOTS" -gt "$CORE_CAP" ] && { SLOTS="$CORE_CAP"; BIND="cores"; }
 
-echo "[do-fleet] ranking… (top $TOP candidates, $SLOTS slots, $( $GO && echo LIVE || echo DRY-RUN ))"
+# …and by MEMORY, which is the constraint that actually bites.
+#
+# Cores were never the limit. On a 10-core/24GB box `cores - 2` authorised 8
+# concurrent worktrees; 8 cycles is ~16GB of gates before the editors, the
+# sessions and the OS get a byte. The box reached load 57 and filled 16GB of
+# swap with fewer than that. gate_headroom prices a cycle in GB and divides
+# what is actually free, so the fleet shrinks on a loaded box and opens back
+# up on an idle one. It only ever LOWERS the count — --slots and the core cap
+# still hold.
+# shellcheck source=lib/govern.sh
+. "$ROOT/.claude/scripts/lib/govern.sh"
+MEM_CAP=$(gate_headroom)
+[ "$SLOTS" -gt "$MEM_CAP" ] && { SLOTS="$MEM_CAP"; BIND="memory"; }
+
+# A worktree cut before the governor landed runs the OLD unbounded gates, so
+# the cap above is the only thing protecting the box from it. Say so — a silent
+# stale worktree is how this problem returns.
+for _wt in "$WT_BASE"/*/; do
+  [ -f "$_wt/.claude/scripts/lib/govern.sh" ] && continue
+  [ -d "$_wt/.claude/scripts" ] || continue
+  echo "[do-fleet] ⚠ $(basename "$_wt") predates the governor — its gates are unbounded; rebase it onto $BASE_REF"
+done
+
+echo "[do-fleet] ranking… (top $TOP candidates, $SLOTS slots — bound by $BIND; cores≤$CORE_CAP mem≤$MEM_CAP, $( $GO && echo LIVE || echo DRY-RUN ))"
 
 # Load ranked queue into an array (bash 3.2 compatible — no mapfile).
 QUEUE=()
@@ -89,47 +113,82 @@ link_deps() {
   done
 }
 
-# Dynamic file-lock registry: "dir<TAB>slug" lines in a temp file.
-LOCK_FILE="$(mktemp)"
-trap 'rm -f "$LOCK_FILE"' EXIT
-
-claim_dirs() {   # claim_dirs <slug>
-  local n=0
+# Region-claim registry — MACHINE-WIDE, in $GOVERN_DIR/claims (lib/govern.sh).
+#
+# This used to be `LOCK_FILE="$(mktemp)"`: a map private to one fleet process.
+# Nine concurrent fleet invocations held nine private maps, so two fleets could
+# each believe a region was free and both take it — the duplicate-selection
+# collision this registry exists to remove. mkdir(2) is the arbiter; a claim
+# evaporates on its own when the owner dies (kill -0) or its lease expires.
+trap 'claim_release_all' EXIT
+
+claim_dirs() {   # claim_dirs <slug> → 0 = all regions claimed, 1 = lost a race
+  local slug="$1" n=0 d
+  CLAIM_TAKEN=()
   while IFS= read -r d; do
-    [ -n "$d" ] && { printf '%s\t%s\n' "$d" "$1" >> "$LOCK_FILE"; n=$(( n + 1 )); }
-  done < <(deliverable_dirs "$1")
+    [ -n "$d" ] || continue
+    n=$(( n + 1 ))
+    if claim_take "$d" "$slug"; then
+      CLAIM_TAKEN+=("$d")
+    else
+      # Another fleet won this region between has_conflict and here. Give back
+      # everything already taken for this slug — a half-claimed plan would hold
+      # ground it never builds, and only the TTL would ever free it.
+      if [ "${#CLAIM_TAKEN[@]}" -gt 0 ]; then
+        for d in "${CLAIM_TAKEN[@]}"; do claim_release "$d"; done
+      fi
+      echo "  ✗ $slug — lost the race for a region to another fleet, deferring" >&2
+      return 1
+    fi
+  done < <(deliverable_dirs "$slug")
   # An unscoped plan (no deliverables:) declares zero dirs — the conflict filter
   # would treat it as disjoint from EVERYTHING and parallelize it blindly, which
   # clobbers shared trees (one.ie/web). Claim the __ANY__ sentinel so it conflicts
   # with any other plan: it only ever runs as the sole slot. Declare deliverables
   # to make it fleet-safe (see do-rank.py UNSCOPED warning).
-  [ "$n" -eq 0 ] && printf '%s\t%s\n' "__ANY__" "$1" >> "$LOCK_FILE"
+  if [ "$n" -eq 0 ]; then
+    claim_take "__ANY__" "$slug" || { echo "  ✗ $slug — __ANY__ taken, deferring" >&2; return 1; }
+  fi
+  return 0
 }
 
 free_dirs() {    # free_dirs <slug>
-  local tmp; tmp="$(mktemp)"
-  grep -v "	$1$" "$LOCK_FILE" > "$tmp" 2>/dev/null || true
-  mv "$tmp" "$LOCK_FILE"
+  local d
+  while IFS= read -r d; do
+    [ -n "$d" ] && claim_release "$d"
+  done < <(deliverable_dirs "$1")
+  # Only the slug that took the sentinel may give it back. Every claim in this
+  # process shares one pid, so an unguarded release here would let a finishing
+  # SCOPED plan hand back a live UNSCOPED plan's sole-slot guarantee.
+  [ "$(claim_owner "__ANY__" | awk '{print $1}')" = "$1" ] && claim_release "__ANY__"
+  return 0
 }
 
 has_conflict() { # has_conflict <slug> → exit 0 if conflict
+  # Every read below goes through claim_owner/claim_list, which reap a dead or
+  # expired owner before answering — so a fleet that was killed mid-plan never
+  # keeps a region shut.
+  local live_owner
   # A live unscoped plan holds the __ANY__ sentinel → nothing may share with it.
-  any_owner=$(awk -F'\t' '$1=="__ANY__"{print $2; exit}' "$LOCK_FILE" 2>/dev/null || true)
+  any_owner=$(claim_owner "__ANY__" | awk '{print $1}')
   if [ -n "$any_owner" ] && [ "$any_owner" != "$1" ]; then
     echo "  ✗ $1 — unscoped plan $any_owner holds the fleet (sole slot), deferring"
     return 0
   fi
   # This plan is itself unscoped (no deliverable dirs) → only runs alone.
   ndirs=$(deliverable_dirs "$1" | grep -c . || true)
-  if [ "$ndirs" -eq 0 ] && [ -s "$LOCK_FILE" ]; then
-    echo "  ✗ $1 — unscoped (no deliverables:), defers while any plan is live"
-    return 0
+  if [ "$ndirs" -eq 0 ]; then
+    live_owner=$(claim_list | awk -F'\t' -v s="$1" '$2!=s{print $2; exit}')
+    if [ -n "$live_owner" ]; then
+      echo "  ✗ $1 — unscoped (no deliverables:), defers while $live_owner is live"
+      return 0
+    fi
   fi
   while IFS= read -r d; do
     [ -z "$d" ] && continue
-    owner=$(awk -F'\t' -v k="$d" '$1==k{print $2; exit}' "$LOCK_FILE" 2>/dev/null || true)
-    if [ -n "$owner" ]; then
-      echo "  ✗ $1 — dir $d locked by $owner, deferring"
+    owner=$(claim_owner "$d" | awk '{print $1}')
+    if [ -n "$owner" ] && [ "$owner" != "$1" ]; then
+      echo "  ✗ $1 — dir $d claimed by $owner, deferring"
       return 0
     fi
   done < <(deliverable_dirs "$1")
@@ -175,16 +234,22 @@ claim_task() {  # claim_task <slug> → 0 ok to launch, 1 already claimed elsewh
   esac
 }
 
-launch_plan() {  # launch_plan <slug> → echoes PID
+launch_plan() {  # launch_plan <slug> → echoes PID, or nothing if the claim was lost
   local slug="$1"
   local wt="$WT_BASE/$slug"
+  # Claim the ground FIRST. has_conflict only read the registry; between that
+  # read and now another fleet may have deposited. mkdir decides, and a loser
+  # reroutes to the next candidate rather than waiting.
+  claim_dirs "$slug" || return 1
   if [ ! -d "$wt" ]; then
     git -C "$ROOT" worktree add -b "do/$slug" "$wt" "$BASE_REF" >/dev/null 2>&1 \
       || git -C "$ROOT" worktree add "$wt" "do/$slug" >/dev/null 2>&1 || true
   fi
-  link_deps "$wt"
-  claim_dirs "$slug"
-  echo "  ▶ launching $slug (log: $wt/.fleet.log)"
+  # stdout of this function IS the pid — fill_slots captures it. Chatter must go
+  # to stderr or the "pid" is a multi-line string and every kill -0 on it fails,
+  # which reads as "slot finished" the instant it launches.
+  link_deps "$wt" >&2
+  echo "  ▶ launching $slug (log: $wt/.fleet.log)" >&2
   # Direct exec (no subshell wrapper) so $! is the do-auto.sh bash process itself.
   # A ( cmd ) & subshell forks cmd and exits immediately — the fleet would track the
   # subshell PID, which dies before cmd finishes, misreporting the slot as done.
@@ -196,7 +261,15 @@ launch_plan() {  # launch_plan <slug> → echoes PID
 if ! $GO; then
   echo ""
   echo "  DRY-RUN — first $SLOTS disjoint plans from queue:"
-  CLAIMED_DRY=""
+  # Seed the preview from the LIVE machine-wide registry (read-only — a dry run
+  # never deposits). Otherwise the preview says "✓ foo" for a region another
+  # fleet is already inside, which is exactly the lie the registry removes.
+  # claim_list emits region<TAB>slug<TAB>pid and reaps dead owners as it walks,
+  # so the preview inherits evaporation for free.
+  CLAIMED_DRY="$(claim_list)"
+  # Keep "" meaning "nothing claimed" — the unscoped-plan check below tests -n.
+  if [ -n "$CLAIMED_DRY" ]; then CLAIMED_DRY="$CLAIMED_DRY
+"; fi
   COUNT=0
   for slug in "${QUEUE[@]}"; do
     [ "$COUNT" -ge "$SLOTS" ] && break
@@ -309,7 +382,12 @@ fill_slots() {
       DONE_SLUGS+=("$NEXT_SLUG")
       continue
     fi
-    local pid; pid=$(launch_plan "$NEXT_SLUG")
+    local pid=""
+    # A lost region race is not an error — reroute to the next candidate.
+    if ! pid=$(launch_plan "$NEXT_SLUG") || [ -z "$pid" ]; then
+      DONE_SLUGS+=("$NEXT_SLUG")
+      continue
+    fi
     SLOT_PIDS+=("$pid")
     SLOT_SLUGS+=("$NEXT_SLUG")
     SLOT_START+=("$(date +%s)")

package/scripts/do-folder.sh

@@ -36,7 +36,11 @@ fi
 
 # 3. longest-prefix match against known buildable folders (one.ie/web BEFORE one.ie).
 #    bash 3.2-safe (macOS): no associative arrays — dedup a newline list with sort -u.
-folders="one.ie/web one.ie packages agents api schema sync backup"
+# channels/ and pay/ were ABSENT until 2026-09-03, and they are two of the five
+# deployed services. A changed set touching only them answered "unmapped", which
+# do-w4-gates then dropped on the floor -- no verify row at all, in either lane.
+# pay/backend before pay, one.ie/web before one.ie: longest prefix must win.
+folders="one.ie/web one.ie packages agents api schema sync backup channels pay/backend pay"
 matched=""
 for p in "${paths[@]}"; do
   for f in $folders; do
@@ -56,7 +60,12 @@ printf '%s\n' "$matched" | while IFS= read -r f; do
   pj="$ROOT/$f/package.json"
   verify="null"; build="null"
   if [ -f "$pj" ]; then
+    # Dev lane by default: verify:fast (tsc + related tests + pinned gates).
+    # FULL_VERIFY=1 (set by /close and ./deploy) asks for the whole suite.
     jq -e '.scripts.verify' "$pj" >/dev/null 2>&1 && verify='"bun run verify"'
+    if [ "${FULL_VERIFY:-0}" != "1" ] && jq -e '.scripts["verify:fast"]' "$pj" >/dev/null 2>&1; then
+      verify='"bun run verify:fast"'
+    fi
     jq -e '.scripts.build'  "$pj" >/dev/null 2>&1 && build='"bun run build"'
   fi
   printf '{"folder":"%s","verify":%s,"build":%s,"doc_only":false}\n' "$f" "$verify" "$build"