@tekyzinc/gsd-t 4.6.10 → 4.7.10

package/CHANGELOG.md

@@ -2,6 +2,35 @@
 
 All notable changes to GSD-T are documented here. Updated with each release.
 
+## [4.7.10] - 2026-06-22 (M90 — The Unproven-Assumption Doctrine — minor)
+
+### Added — a self-governing doctrine that stops the system building on unproven assumptions
+
+M90 makes "don't act on belief; if it's not grounded in knowledge or research, research first" an enforced framework capability rather than advisory prose. The core principle (sourced at discuss: self-introspection is unreliable, *worse* on RLHF models — arxiv 2306.13063, 2310.01798) is **externalize + force, never introspect** — deterministic triggers + external checks, fail-closed on uncertainty.
+
+Three deterministic mechanisms, wired fail-closed into the debug/execute/quick/phase/verify workflows (§4 R-FAIL-1/2/3):
+
+- **Factual classifier** (`bin/gsd-t-research-gate.cjs`, absorbed from M89): routes a knowledge-gap to grep (internal) vs. research+cite (external); the INTERNAL decision enumerates no open category (never guess-internal). Premise-corrected during build (the partition's "vendor list causes a silent-miss" was false on disk — the list is kept as an `external→web` upgrade). New time-anchored override (R-FACT-3): a fast-moving lib/API/version or "current best practice" gap is researched regardless of confidence.
+- **Loop-ledger non-convergence halt** (`bin/gsd-t-loop-ledger.cjs`): a debug loop on the SAME symptom (keyed on the symptom, not the changing file — so variant-spawning whack-a-mole is caught, R-LOOP-1) HARD-HALTS with a premise-re-examination directive instead of patching further. Detection ≠ resolution (the halt persists until a genuine re-examination clears it), per-milestone scoped (one milestone's halt can't brick another's verify), full-reset on re-examination, plain-object-validated, atomic write.
+- **Architectural-assumption trigger** (`bin/gsd-t-architectural-trigger.cjs`): a divergence-sampling + extend-existing-code trigger that flags building on an unproven approach. Shipped EXPERIMENTAL+MEASURED; the spike/adversary response-mode enforcement (R-FAIL-2) is honestly DECLARED interface-only this milestone (no live producer yet — backlog #42 wires it).
+
+Doctrine contract `unproven-assumption-doctrine-contract.md` v1.0.0 STABLE (absorbs `auto-research-contract.md` v1.3.3) + §6 guard map + self-obedience lints. CLAUDE-global Research Policy upgraded from advisory prose to the doctrine.
+
+### Process note
+
+The milestone obeyed its own doctrine: plan-hardening caught M90 about to build a domain on an unproven premise (re-verified on disk, re-scoped), and a 9-round verify sequence hardened M90's OWN gate-wiring lifecycle to ground (hollow-gate → fail-open → no-producer → lifecycle flaws → vacuous-gate → array-type-confusion → cross-milestone-brick → symptom-keying). Suite 1998 / 1994 pass / 0 fail / 4 skip. Deferred to backlog: #42 (live spike-feasibility producer + sink fail-closed), #43 (AWS/S3 classifier edge), 3 documented code-review nits.
+
+## [4.6.11] - 2026-06-16 (Output Style — six named conciseness tics + backlog #33 — patch)
+
+### Changed — tightened the CONCISE Output Style rule with six named anti-patterns
+
+User feedback that replies stayed wordy despite the existing CONCISE rule. Extracted six specific tics — three from live examples, three from the user's own wordy→concise rewrites (binvoice `Wordy Example 1/2.txt`, a labeled before/after set where ~70 wordy lines collapsed to ~25 with identical information) — and added a litmus test. Synced the template to the live global CLAUDE.md (blocks kept identical — ripple invariant). Also added backlog #33 from the binvoice FB-modal debug-loop retrospective.
+
+- `templates/CLAUDE-global.md`: Output Style block gains six rules — no process narration, no answer sandwich, no affirmation throat-clearing, no honesty theater, a table replaces its prose (never repeats it), ask once — plus a litmus test ("delete any sentence that survives deletion without info loss").
+- `.gsd-t/backlog.md`: #33 firing debug-cycle circuit-breaker + repro-fixture-on-regression + anchor-last scraping stack rule (completes #31/TD-294 from the loop-governance side).
+
+Behavioral/doc-only — no test changes. Suite: 1603/1607 pass, 0 fail.
+
 ## [4.6.10] - 2026-06-15 (Installer wiring for status line + low-context cue — minor)
 
 ### Added — installer now copies and wires the status line + ctx-cue Stop hook

package/README.md

@@ -1,6 +1,6 @@
 # GSD-T: Contract-Driven Development for Claude Code
 
-**v4.5.10** - A methodology for reliable, parallelizable development using Claude Code with optional Agent Teams support.
+**v4.7.10** - A methodology for reliable, parallelizable development using Claude Code with optional Agent Teams support.
 
 **Eliminates context rot** — task-level fresh dispatch (one subagent per task, ~10-20% context each) means compaction never triggers.
 **Compaction-proof debug loops** — `gsd-t headless --debug-loop` runs test-fix-retest cycles as separate `claude -p` sessions. A JSONL debug ledger persists all hypothesis/fix/learning history across fresh sessions. Anti-repetition preamble injection prevents retrying failed hypotheses. Escalation tiers (sonnet → opus → human) and a hard iteration ceiling enforced externally.

package/bin/gsd-t-architectural-trigger.cjs

@@ -0,0 +1,523 @@
+"use strict";
+
+/**
+ * gsd-t-architectural-trigger.cjs — M90 D1 (v1.0.0)
+ *
+ * Deterministic architectural-assumption trigger module.
+ *
+ * TWO fire paths:
+ *
+ *   R-ARCH-1 — DIVERGENCE SAMPLING (experimental + measured)
+ *     Given N≥2 fresh-context answers to the SAME approach question, computes a
+ *     deterministic divergence measure (normalized type-token ratio variance across
+ *     answer n-gram fingerprints). When divergence exceeds DIVERGENCE_THRESHOLD, fires
+ *     with reason "divergence-sampling:high-variance". Sourced from ClarifyGPT
+ *     consistency-sampling (https://arxiv.org/pdf/2310.10996): divergence across
+ *     fresh-context samples ⇒ a load-bearing unproven assumption exists.
+ *     NOTE: This path is EXPERIMENTAL+MEASURED (competition-arm only). Self-MoA
+ *     (one model, temperature-varied) may not diverge like the fresh-context saga
+ *     cases the threshold was tuned on. That mismatch is RECORDED by the
+ *     instrumentation sink; the path never claims it works.
+ *
+ *   R-ARCH-2 — PROTOCOL CLASS (unconditional on extend-existing-code signal)
+ *     Extending existing code is itself an approach decision — the trigger fires
+ *     unconditionally on an extend signal (no confidence/threshold gate). Signal
+ *     format: { type: "extend-existing-code", context, basis }.
+ *     Signal is COMPUTED at runtime from task/scope inputs that already exist
+ *     (a task whose **Touches** lists an EXISTING file ⇒ extend-class). The
+ *     D-CONTRACT domain (D4-T4) computes this from brief/task inputs.
+ *     Reason literal: "protocol-class:extend-existing" (distinguishable from R-ARCH-1).
+ *
+ * RESPONSE INTERFACE (R-ARCH-3..6) — INTERFACE ONLY; agent() wiring is D-CONTRACT.
+ *
+ *   mode "spike" — PREFERRED response; attempt an executable spike to prove feasibility.
+ *   mode "adversary-only" — fallback when spike is infeasible or fails.
+ *   R-ARCH-4: spike-fail → STOP directive in envelope; agent cannot proceed.
+ *   R-ARCH-5: spike-infeasible → logged skip + adversary MANDATORY.
+ *   R-ARCH-6: proven-by-adversary-only → flag surfaced for verify gate.
+ *
+ * INSTRUMENTATION (no-claim invariant):
+ *   Every fire emits a JSONL record to the measurement sink. The module NEVER asserts
+ *   it works — it emits data only (fire-rate, catch-quality fields). No self-efficacy
+ *   claim in any envelope field.
+ *
+ * HOUSE STYLE:
+ *   { ok: true, ... }    on success
+ *   { ok: false, error } on bad input
+ *   Deterministic: identical input → byte-identical envelope.
+ *   Bad input → { ok:false, error } + non-zero CLI exit.
+ *   Node built-ins only; zero new runtime deps; sync APIs.
+ *
+ * CONTRACT: .gsd-t/contracts/m90-doctrine-mechanisms-contract.md §2 v1.0.0
+ */
+
+const fs = require("fs");
+const path = require("path");
+const os = require("os");
+
+// ---------------------------------------------------------------------------
+// Constants (DECLARED — not back-fit per fixture)
+// ---------------------------------------------------------------------------
+
+/**
+ * Divergence threshold for the shared-core-term fraction metric.
+ * DECLARED CONSTANT — the divergence measure is a fixed formula and this threshold
+ * is a declared constant, not back-fit to any fixture.
+ *
+ * Formula: score = 1 - (sharedCoreTerms / allUniqueTerms)
+ *   - sharedCoreTerms = non-stopword tokens (≥4 chars) appearing in ≥50% of N answers
+ *   - allUniqueTerms = total unique non-stopword tokens across all answers
+ *
+ * Basis: ClarifyGPT (arXiv:2310.10996) consistency-sampling — across N independent
+ * fresh-context samples, low shared-core-term fraction (high unique-term proportion)
+ * indicates each answer introduced different technical mechanisms, the proxy for
+ * "an unproven architectural premise exists." A HIGH shared-core fraction means
+ * all answers converged on the same technical strategy.
+ *
+ * Threshold 0.80 was chosen as the midpoint of the confirmed gap in the TUNED corpus:
+ *   divergent rows all score ≥0.89; convergent rows all score ≤0.70.
+ * Gap = 0.19; midpoint ≈ 0.79. Set to 0.80 (conservative toward precision over recall).
+ * NOT tuned to held-out rows.
+ */
+const DIVERGENCE_THRESHOLD = 0.80;
+
+/**
+ * Minimum N for divergence sampling. N<2 cannot produce a meaningful comparison.
+ */
+const MIN_ANSWERS = 2;
+
+/**
+ * Stopwords excluded from token analysis.
+ * Common English function words plus broad technical/generic terms that appear in
+ * BOTH divergent and convergent contexts equally (non-discriminative).
+ */
+const STOPWORDS = new Set([
+  "the", "a", "an", "is", "are", "was", "were", "be", "been", "being",
+  "have", "has", "had", "do", "does", "did", "will", "would", "could",
+  "should", "may", "might", "must", "shall", "to", "of", "in", "on",
+  "at", "by", "for", "with", "from", "as", "or", "and", "but", "not",
+  "it", "its", "this", "that", "these", "those", "we", "our", "you",
+  "your", "they", "them", "i", "me", "my", "he", "she", "him", "her",
+  "which", "what", "how", "when", "where", "if", "so", "also", "just",
+  "then", "than", "there", "here", "same", "all", "any", "each", "can",
+  "into", "via", "no", "only", "need", "using", "use", "call", "make",
+  "ensure", "before", "after", "pass", "get", "set", "both", "up",
+  "out", "above", "below", "end", "result", "object", "string",
+  "function", "value", "input", "output", "code", "signal",
+]);
+
+/**
+ * Measurement sink file location. Relative to the project root (cwd at runtime).
+ * Kept in .gsd-t/metrics/ alongside other measurement sinks.
+ */
+const SINK_RELATIVE_PATH = path.join(".gsd-t", "metrics", "arch-trigger-events.jsonl");
+
+// ---------------------------------------------------------------------------
+// Divergence measure (R-ARCH-1)
+// ---------------------------------------------------------------------------
+
+/**
+ * Tokenize an answer string into a SET of unique lowercase key terms.
+ * Filters out stopwords and short tokens (< 4 chars) to focus on
+ * substantive technical vocabulary.
+ *
+ * Deterministic: same string → same token set.
+ *
+ * @param {string} text
+ * @returns {Set<string>}
+ */
+function tokenize(text) {
+  const tokens = (text.match(/[a-z][a-z0-9_.-]*/gi) || [])
+    .map((t) => t.toLowerCase())
+    .filter((t) => t.length >= 4 && !STOPWORDS.has(t));
+  return new Set(tokens);
+}
+
+/**
+ * Given N fresh-context answers to the same approach question, compute a
+ * deterministic divergence score using the shared-core-term fraction metric.
+ *
+ * THE FIXED, DOCUMENTED FORMULA:
+ *   1. Tokenize each answer into a set of unique non-stopword terms (≥4 chars).
+ *   2. For each unique term across all answers, count how many answers contain it.
+ *   3. sharedCoreTerms = count of terms appearing in ≥ ceil(N * 0.5) answers.
+ *      (i.e., appearing in at least half of all N answers)
+ *   4. allUniqueTerms = total count of distinct terms across all N answers.
+ *   5. score = 1 - (sharedCoreTerms / allUniqueTerms)
+ *
+ * Interpretation:
+ *   - score ≈ 0: most terms are shared → answers agree on same strategy → CONVERGENT
+ *   - score ≈ 1: few terms are shared → each answer introduced different mechanisms → DIVERGENT
+ *
+ * Basis: divergent answers each name a different technical mechanism (different
+ * key tokens). Convergent answers share the same core technical vocabulary because
+ * they all describe the same strategy. The shared-term fraction measures how much
+ * vocabulary each answer contributes that was NOT in the others.
+ *
+ * Threshold DIVERGENCE_THRESHOLD = 0.80. Gap between corpus classes:
+ *   divergent ≥ 0.89, convergent ≤ 0.70, gap = 0.19.
+ *
+ * Deterministic: identical inputs (same strings in same order) → byte-identical score.
+ * allUniqueTerms = 0 → returns 0 (degenerate all-empty input; no fire).
+ *
+ * @param {string[]} answers - N≥2 fresh-context answers to the same question.
+ * @returns {number} Divergence score in [0, 1].
+ */
+function computeDivergenceScore(answers) {
+  const tokenSets = answers.map((a) => tokenize(a));
+  const termCounts = new Map();
+
+  for (const set of tokenSets) {
+    for (const term of set) {
+      termCounts.set(term, (termCounts.get(term) || 0) + 1);
+    }
+  }
+
+  if (termCounts.size === 0) return 0;
+
+  const N = answers.length;
+  const sharedThreshold = Math.max(2, Math.ceil(N * 0.5));
+
+  let sharedCount = 0;
+  for (const count of termCounts.values()) {
+    if (count >= sharedThreshold) sharedCount++;
+  }
+
+  return 1 - sharedCount / termCounts.size;
+}
+
+// ---------------------------------------------------------------------------
+// Instrumentation sink (no-claim invariant)
+// ---------------------------------------------------------------------------
+
+/**
+ * Emit a structured instrumentation record to the measurement sink.
+ *
+ * NEVER adds a self-efficacy claim. Records facts ONLY:
+ *   - fired (boolean)
+ *   - basis (string | null)
+ *   - mode (string — the response mode selected)
+ *   - divergenceScore (number | null — for R-ARCH-1 path)
+ *   - firePath ("divergence-sampling" | "protocol-class" | null)
+ *
+ * Wall-clock timestamp is NOT in the asserted envelope shape (the trigger envelope
+ * is deterministic). The sink record carries an ISO timestamp for observability but
+ * it is NOT in the trigger envelope returned to callers.
+ *
+ * Emission is best-effort — if the sink directory does not exist or the write fails,
+ * the error is SWALLOWED (the trigger result is never blocked by instrumentation).
+ *
+ * @param {object} record
+ */
+function emitInstrumentationRecord(record) {
+  try {
+    const sinkPath = path.resolve(process.cwd(), SINK_RELATIVE_PATH);
+    const dir = path.dirname(sinkPath);
+    if (!fs.existsSync(dir)) {
+      fs.mkdirSync(dir, { recursive: true });
+    }
+    // ISO timestamp for observability (not in the asserted envelope)
+    const entry = Object.assign({}, record, { _emittedAt: new Date().toISOString() });
+    fs.appendFileSync(sinkPath, JSON.stringify(entry) + os.EOL, "utf8");
+  } catch (_) {
+    // Instrumentation failure NEVER blocks the trigger result.
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Response mode resolver (R-ARCH-3..6)
+// ---------------------------------------------------------------------------
+
+/**
+ * Resolve the response mode flags given the spike feasibility / spike result inputs.
+ *
+ * @param {object} opts
+ * @param {boolean} [opts.spikeFeasible=true]  - Can an executable spike be run?
+ * @param {boolean} [opts.spikePassed]         - Did the spike pass? (ignored if spikeFeasible=false)
+ * @returns {object} Response mode envelope fragment.
+ */
+function resolveResponseMode({ spikeFeasible = true, spikePassed } = {}) {
+  if (!spikeFeasible) {
+    // R-ARCH-5: spike infeasible → adversary-only + adversaryMandatory + logged skip
+    return {
+      mode: "adversary-only",
+      adversaryMandatory: true,
+      spikeSkipReason: "spike-infeasible: executable spike cannot be constructed for this premise",
+      provenByAdversaryOnly: true,
+      stopDirective: false,
+    };
+  }
+
+  if (spikePassed === false) {
+    // R-ARCH-4: spike ran and FAILED → STOP; agent cannot proceed
+    return {
+      mode: "adversary-only",
+      adversaryMandatory: true,
+      provenByAdversaryOnly: false,
+      stopDirective: true,
+      stopReason: "spike-fail: the executable spike failed — the premise is unproven; STOP and re-examine",
+    };
+  }
+
+  if (spikePassed === true) {
+    // Spike passed → prefer spike, adversary still RECOMMENDED but not mandatory
+    return {
+      mode: "spike",
+      adversaryMandatory: false,
+      provenByAdversaryOnly: false,
+      stopDirective: false,
+    };
+  }
+
+  // Default: spike PREFERRED (no spike result provided yet)
+  return {
+    mode: "spike",
+    adversaryMandatory: false,
+    provenByAdversaryOnly: false,
+    stopDirective: false,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Main trigger function
+// ---------------------------------------------------------------------------
+
+/**
+ * Validate shared bad-input conditions.
+ * Returns { ok:false, error } if invalid, else null.
+ *
+ * @param {*} input
+ * @returns {{ ok:false, error:string }|null}
+ */
+function validateInput(input) {
+  if (input === null || input === undefined) {
+    return { ok: false, error: "input must be a non-null object" };
+  }
+  if (typeof input !== "object" || Array.isArray(input)) {
+    return { ok: false, error: "input must be a plain object" };
+  }
+  return null;
+}
+
+/**
+ * R-ARCH-1 divergence-sampling trigger.
+ *
+ * @param {object} input
+ * @param {string[]} input.answers - N≥2 fresh-context answers to the same question.
+ * @param {string}   input.basis   - The approach/premise being evaluated.
+ * @param {object}   [input.responseOpts] - Optional response mode options (see resolveResponseMode).
+ * @returns {object} House-style envelope.
+ */
+function triggerDivergenceSampling(input) {
+  const badInput = validateInput(input);
+  if (badInput) return badInput;
+
+  const { answers, basis, responseOpts = {} } = input;
+
+  if (!Array.isArray(answers)) {
+    return { ok: false, error: "input.answers must be an array of strings" };
+  }
+  if (answers.length < MIN_ANSWERS) {
+    return {
+      ok: false,
+      error: `input.answers must contain at least ${MIN_ANSWERS} elements; got ${answers.length}`,
+    };
+  }
+  for (let i = 0; i < answers.length; i++) {
+    if (typeof answers[i] !== "string") {
+      return { ok: false, error: `input.answers[${i}] must be a string` };
+    }
+  }
+  if (!basis || typeof basis !== "string" || !basis.trim()) {
+    return { ok: false, error: "input.basis must be a non-empty string" };
+  }
+
+  const trimmedBasis = basis.trim();
+  const score = computeDivergenceScore(answers);
+  const fired = score >= DIVERGENCE_THRESHOLD;
+
+  const responseMode = resolveResponseMode(responseOpts);
+
+  const envelope = {
+    ok: true,
+    firePath: "divergence-sampling",
+    fired,
+    basis: trimmedBasis,
+    reason: fired
+      ? "divergence-sampling:high-variance"
+      : "divergence-sampling:low-variance",
+    divergenceScore: score,
+    divergenceThreshold: DIVERGENCE_THRESHOLD,
+    experimental: true, // NEVER omit: this path is experimental+measured, never proven
+    ...(fired ? responseMode : {}),
+  };
+
+  // Emit instrumentation (no-claim invariant: emitting a record, not asserting reliability).
+  // provenByAdversaryOnly + mode are PERSISTED so the verify R-FAIL-2 gate (§4) — which scans this
+  // sink for provenByAdversaryOnly===true — has the field it checks; without it the gate is vacuous
+  // (count always 0, can never fire), the exact hollow-gate failure M90 exists to prevent.
+  emitInstrumentationRecord({
+    firePath: "divergence-sampling",
+    fired,
+    basis: trimmedBasis,
+    mode: fired ? responseMode.mode : null,
+    provenByAdversaryOnly: fired ? !!responseMode.provenByAdversaryOnly : false,
+    divergenceScore: score,
+  });
+
+  return envelope;
+}
+
+/**
+ * R-ARCH-2 protocol-class trigger on extend-existing-code signal.
+ *
+ * @param {object} input
+ * @param {string} input.context - Description of the existing code being extended.
+ * @param {string} input.basis   - The approach/premise being evaluated.
+ * @param {object} [input.responseOpts] - Optional response mode options.
+ * @returns {object} House-style envelope.
+ */
+function triggerExtendExistingCode(input) {
+  const badInput = validateInput(input);
+  if (badInput) return badInput;
+
+  const { context, basis, responseOpts = {} } = input;
+
+  if (!context || typeof context !== "string" || !context.trim()) {
+    return { ok: false, error: "input.context must be a non-empty string" };
+  }
+  if (!basis || typeof basis !== "string" || !basis.trim()) {
+    return { ok: false, error: "input.basis must be a non-empty string" };
+  }
+
+  const trimmedBasis = basis.trim();
+  const responseMode = resolveResponseMode(responseOpts);
+
+  const envelope = {
+    ok: true,
+    firePath: "protocol-class",
+    fired: true, // ALWAYS fires — unconditional on this path (R-ARCH-2)
+    basis: trimmedBasis,
+    reason: "protocol-class:extend-existing", // DISTINCT from divergence path reason
+    context: context.trim(),
+    ...responseMode,
+  };
+
+  // Emit instrumentation. provenByAdversaryOnly + mode PERSISTED for the verify R-FAIL-2 gate (§4).
+  emitInstrumentationRecord({
+    firePath: "protocol-class",
+    fired: true,
+    basis: trimmedBasis,
+    mode: responseMode.mode,
+    provenByAdversaryOnly: !!responseMode.provenByAdversaryOnly,
+    divergenceScore: null,
+  });
+
+  return envelope;
+}
+
+/**
+ * Top-level resolver: dispatch to the appropriate fire path.
+ *
+ * The `type` field in input selects the path:
+ *   "divergence-sampling"   → R-ARCH-1 (input.answers[] required)
+ *   "extend-existing-code"  → R-ARCH-2 (input.context required)
+ *
+ * Bad input (missing type, unknown type, empty/whitespace/non-string fields) →
+ * { ok:false, error } + non-zero CLI exit.
+ *
+ * @param {object} input
+ * @returns {object} House-style envelope.
+ */
+function resolve(input) {
+  const badInput = validateInput(input);
+  if (badInput) return badInput;
+
+  const { type } = input;
+
+  if (!type || typeof type !== "string" || !type.trim()) {
+    return { ok: false, error: 'input.type must be "divergence-sampling" or "extend-existing-code"' };
+  }
+
+  switch (type.trim()) {
+    case "divergence-sampling":
+      return triggerDivergenceSampling(input);
+    case "extend-existing-code":
+      return triggerExtendExistingCode(input);
+    default:
+      return {
+        ok: false,
+        error: `Unknown type "${type}". Expected "divergence-sampling" or "extend-existing-code"`,
+      };
+  }
+}
+
+// ---------------------------------------------------------------------------
+// CLI entry point
+// ---------------------------------------------------------------------------
+//
+// Usage:
+//   gsd-t-architectural-trigger trigger '{"type":"divergence-sampling","answers":[...],"basis":"..."}'
+//   gsd-t-architectural-trigger trigger '{"type":"extend-existing-code","context":"...","basis":"..."}'
+//
+// Bad input → { ok:false, error } + non-zero exit.
+
+if (require.main === module) {
+  const args = process.argv.slice(2);
+
+  function emit(obj) {
+    process.stdout.write(JSON.stringify(obj) + "\n");
+  }
+  function emitError(msg, exitCode) {
+    emit({ ok: false, error: msg });
+    process.exit(exitCode);
+  }
+
+  const subcommand = args[0];
+  if (!subcommand || subcommand.trim() === "") {
+    emitError(
+      'Missing subcommand. Usage: gsd-t-architectural-trigger trigger \'<JSON input>\'',
+      64
+    );
+  }
+
+  if (subcommand !== "trigger") {
+    emitError(
+      `Unknown subcommand "${subcommand}". Usage: gsd-t-architectural-trigger trigger '<JSON input>'`,
+      64
+    );
+  }
+
+  const rawInput = args[1];
+  if (rawInput === undefined || rawInput === null || rawInput.trim() === "") {
+    emitError(
+      "Missing JSON input. Usage: gsd-t-architectural-trigger trigger '<JSON input>'",
+      64
+    );
+  }
+
+  let parsed;
+  try {
+    parsed = JSON.parse(rawInput);
+  } catch (e) {
+    emitError(`Malformed JSON input: ${e.message}`, 64);
+  }
+
+  const result = resolve(parsed);
+  emit(result);
+  if (!result.ok) process.exit(1);
+  // Success → exit 0
+}
+
+// ---------------------------------------------------------------------------
+// Module exports (frozen §2 signature — DO NOT CHANGE after Wave 1 closes)
+// ---------------------------------------------------------------------------
+
+module.exports = {
+  resolve,
+  triggerDivergenceSampling,
+  triggerExtendExistingCode,
+  resolveResponseMode,
+  computeDivergenceScore,
+  DIVERGENCE_THRESHOLD,
+};