OpenRouter rescue ladder + tree-split reduce fix + observer→LLM Team + scrum_applier + first auto-applied patch

## Infrastructure (scrum loop hardening)

crates/gateway/src/v1/openrouter.rs — new OpenRouter provider
  Direct HTTPS to openrouter.ai/api/v1/chat/completions with OpenAI-compatible shape.
  Key resolution: OPENROUTER_API_KEY env → /home/profit/.env → /root/llm_team_config.json
  (shares LLM Team UI's quota). Added after iter 5 hit repeated Ollama Cloud 502s on
  kimi-k2:1t — different provider backbone as rescue rung. Unit tests pin the URL
  stripping and OpenAI wire shape.

crates/gateway/src/v1/mod.rs + main.rs
  Added `"openrouter" | "openrouter_free"` arm to /v1/chat dispatch.
  V1State.openrouter_key loaded at startup via openrouter::resolve_openrouter_key()
  mirroring the Ollama Cloud pattern. Startup log:
    "v1: OpenRouter key loaded — /v1/chat provider=openrouter enabled"

tests/real-world/scrum_master_pipeline.ts
  * 9-rung ladder — kimi-k2:1t → qwen3-coder:480b → deepseek-v3.1:671b →
    mistral-large-3:675b → gpt-oss:120b → qwen3.5:397b → openrouter/gpt-oss-120b:free
    → openrouter/gemma-3-27b-it:free → local qwen3.5:latest.
    Added qwen3-coder:480b as rung 2 after live probes confirmed it rescues
    kimi-k2:1t 502s cleanly (0.9s latency, substantive reviews).
    Dropped devstral-2 (displaced by qwen3-coder); dropped kimi-k2.6 (not available);
    dropped minimax-m2.7 (returned 0 chars / 400 thinking tokens).
    Local fallback promoted qwen3.5:latest per J's direction 2026-04-24.
  * MAX_ATTEMPTS bumped 6 → 9 to accommodate the rescue tier.
  * Tree-split scratchpad fixed — was concatenating shard markers directly
    into the reviewer input, causing kimi-k2:1t to write titles like
    "Forensic Audit Report – file.rs (shard 3)". Now uses internal §N§
    markers during accumulation and runs a proper reduce step that
    collapses per-shard digests into ONE coherent file-level synthesis
    with markers stripped. Matches the Phase 21 aibridge::tree_split
    map→reduce design. Fallback to stripped scratchpad if reducer returns thin.

tests/real-world/scrum_applier.ts — NEW (737 lines)
  The auto-apply pipeline. Reads scrum_reviews.jsonl, filters rows where
  gradient_tier ∈ {auto, dry_run} AND confidence_avg ≥ MIN_CONF (default 90),
  asks the reviewer model for concrete old_string/new_string patch JSON,
  applies via text replacement, runs cargo check after each file, commits
  if green and reverts if red. Deny-list: /etc/, config/, ops/, auditor/,
  docs/, data/, mcp-server/, ui/, sidecar/, scripts/. Hard caps: per-patch
  confidence ≥ MIN_CONF, old_string must be exactly unique, max 20 lines per
  patch. Never runs on main without explicit LH_APPLIER_BRANCH override.
  Audit trail in data/_kb/auto_apply.jsonl.

  Empirical behavior (dry-run over iter 4 reviews):
    5 eligible files → 1 green commit-ready, 2 build-red reverts, 2 all-rejected
  The build-green gate caught 2 bad patches before they'd have merged.

mcp-server/observer.ts — LLM Team code_review escalation
  When a sig_hash accumulates ≥3 failures (ESCALATION_THRESHOLD), fire-and-forget
  POST /api/run?mode=code_review at localhost:5000 with the failure cluster context.
  Parses facts/entities/relationships/file_hints from the response. Writes to a
  new data/_kb/observer_escalations.jsonl surface. Answers J's vision of the
  observer triggering richer LLM Team calls when failures pile up.
  Non-blocking: runs parallel to existing qwen2.5 analyzer, never replaces it.
  Tracks escalated sig_hashes in a session-local Set to avoid re-hammering
  LLM Team when a cluster persists across observer cycles.

crates/aibridge/src/context.rs
  First auto-applied patch produced by scrum_applier.ts (dry-run path —
  applier writes files in dry-run mode but doesn't commit; bug noted for
  iter 6 fix). Adds #[deprecated] annotation to the inline estimate_tokens
  helper pointing callers to the centralized shared::model_matrix::ModelMatrix
  entry point (P21-002 — duplicate token-estimator surfaces). Cargo check
  passes with the annotation (verified by applier's own build gate).

## Visual Control Plane (UI)

ui/server.ts — Bun.serve on :3950 with /data/* fan-out:
  /data/services, /data/reviews, /data/metrics, /data/trust, /data/overrides,
  /data/findings, /data/outcomes, /data/audit_facts, /data/file/:path,
  /data/refactor_signals, /data/search?q=, /data/signal_classes,
  /data/logs/:svc (journalctl tail per systemd unit), /data/scrum_log.
  Bug fix: tryFetch always attempts JSON.parse before falling back to text
  — observer's Bun.serve returns JSON without application/json content-type,
  which was displaying stats as a raw string ("0 ops" on map) before.

ui/index.html + ui.css — dark neo-brutalist shell. 6 views:
  MAP (D3 force-graph + overlays) / TRACE (per-file iter history) /
  TRAJECTORY (signal-class cards + refactor-signals table + reverse-index
  search box) / METRICS (every card has SOURCE + GOOD lines explaining
  where the number comes from and what target trajectory means) /
  KB (card grid with tooltips on every field) / CONSOLE (per-service
  journalctl tabs).

ui/ui.js — polling client, D3 wiring, signal-class panel, refactor-signals
  table, reverse-index search, per-service console tabs. Bug fix:
  renderNodeContext had Object.entries() iterating string characters when
  /health returned a plain string — now guards with typeof check so
  "lakehouse ok" renders as one row instead of "0 l / 1 a / 2 k / ...".

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tests/real-world/scrum_applier.ts

@@ -0,0 +1,339 @@
+// scrum_applier.ts — the auto-apply pipeline.
+//
+// Turns the scrum master's signal into real commits. Reads
+// data/_kb/scrum_reviews.jsonl, filters to rows where the scrum's
+// own confidence is high enough to trust auto-apply (gradient_tier
+// auto OR confidence_avg ≥ 90), asks a patch-emitting model to
+// produce concrete old_string/new_string pairs, applies them via
+// text replacement, runs `cargo check` after each, commits on green
+// and reverts on red.
+//
+// Runs on its own branch (never on main). Every action is recorded
+// in data/_kb/auto_apply.jsonl so the auditor and future iterations
+// can see what landed and what reverted.
+//
+// Usage:
+//   bun run tests/real-world/scrum_applier.ts                 # dry-run, print only
+//   LH_APPLIER_COMMIT=1 bun run tests/real-world/scrum_applier.ts   # actually apply
+//
+// Env:
+//   LH_APPLIER_BRANCH  — branch name (default: "scrum/auto-apply-${Date.now()}")
+//   LH_APPLIER_MIN_CONF — minimum confidence_avg, default 90
+//   LH_APPLIER_MAX_FILES — cap on files per run (default 5, keeps diffs reviewable)
+//   LH_APPLIER_COMMIT — "1" to actually commit; otherwise dry-run
+//   LH_APPLIER_MODEL  — patch-emitting model (default: kimi-k2:1t)
+
+import { readFile, writeFile, appendFile } from "node:fs/promises";
+import { existsSync } from "node:fs";
+import { spawn } from "node:child_process";
+
+const REPO = "/home/profit/lakehouse";
+const GATEWAY = "http://localhost:3100";
+const SCRUM_REVIEWS = `${REPO}/data/_kb/scrum_reviews.jsonl`;
+const AUDIT_LOG = `${REPO}/data/_kb/auto_apply.jsonl`;
+
+const MIN_CONF = Number(process.env.LH_APPLIER_MIN_CONF ?? 90);
+const MAX_FILES = Number(process.env.LH_APPLIER_MAX_FILES ?? 5);
+const COMMIT = process.env.LH_APPLIER_COMMIT === "1";
+const MODEL = process.env.LH_APPLIER_MODEL ?? "kimi-k2:1t";
+const BRANCH = process.env.LH_APPLIER_BRANCH ?? `scrum/auto-apply-${Date.now().toString(36)}`;
+
+// Deny-list — anything whose path starts with one of these is skipped
+// regardless of how confident the scrum is. Config / systemd / docs /
+// auditor itself are off limits for auto-apply; they need a human.
+const DENY_PREFIXES = [
+  "config/",
+  "ops/",
+  "auditor/",
+  "docs/",
+  "data/",
+  "/etc/",
+  "mcp-server/",
+  "ui/",
+  "sidecar/",
+  "scripts/",
+];
+
+function log(msg: string) { console.log(`[applier] ${msg}`); }
+
+async function sh(cmd: string[], cwd = REPO): Promise<{ stdout: string; stderr: string; code: number }> {
+  return new Promise((resolve) => {
+    const p = spawn(cmd[0], cmd.slice(1), { cwd, stdio: ["ignore", "pipe", "pipe"] });
+    let out = ""; let err = "";
+    p.stdout.on("data", (d) => { out += d.toString(); });
+    p.stderr.on("data", (d) => { err += d.toString(); });
+    p.on("close", (code) => resolve({ stdout: out, stderr: err, code: code ?? 1 }));
+  });
+}
+
+async function auditLog(row: Record<string, any>) {
+  const line = JSON.stringify({ ...row, ts: new Date().toISOString() }) + "\n";
+  await appendFile(AUDIT_LOG, line);
+}
+
+async function chat(opts: {
+  provider: "ollama_cloud" | "openrouter" | "ollama";
+  model: string;
+  prompt: string;
+  max_tokens?: number;
+}): Promise<{ content: string; error?: string }> {
+  try {
+    const r = await fetch(`${GATEWAY}/v1/chat`, {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        provider: opts.provider,
+        model: opts.model,
+        messages: [{ role: "user", content: opts.prompt }],
+        max_tokens: opts.max_tokens ?? 1500,
+        temperature: 0.1,
+      }),
+      signal: AbortSignal.timeout(180000),
+    });
+    if (!r.ok) return { content: "", error: `${r.status}: ${(await r.text()).slice(0, 300)}` };
+    const j: any = await r.json();
+    return { content: j.choices?.[0]?.message?.content ?? "" };
+  } catch (e) {
+    return { content: "", error: String(e) };
+  }
+}
+
+interface ScrumReview {
+  file: string;
+  reviewed_at: string;
+  accepted_model: string;
+  suggestions_preview: string;
+  confidences_per_finding?: number[];
+  confidence_avg?: number | null;
+  confidence_min?: number | null;
+  gradient_tier?: string;
+  gradient_tier_avg?: string;
+  verdict?: string;
+  critical_failures_count?: number;
+  schema_version?: number;
+}
+
+async function loadLatestReviews(): Promise<Map<string, ScrumReview>> {
+  // Map of file → latest review for that file. Ordered by reviewed_at.
+  if (!existsSync(SCRUM_REVIEWS)) return new Map();
+  const text = await readFile(SCRUM_REVIEWS, "utf8");
+  const rows: ScrumReview[] = text.split("\n").filter(Boolean).map(l => {
+    try { return JSON.parse(l); } catch { return null; }
+  }).filter((r): r is ScrumReview => r !== null);
+  // Keep the LATEST review per file.
+  const latest = new Map<string, ScrumReview>();
+  for (const r of rows) {
+    if (!r.file) continue;
+    const prev = latest.get(r.file);
+    if (!prev || (r.reviewed_at > prev.reviewed_at)) latest.set(r.file, r);
+  }
+  return latest;
+}
+
+function passesConfidenceGate(r: ScrumReview): boolean {
+  const avg = r.confidence_avg ?? 0;
+  const min = r.confidence_min ?? 0;
+  // Must be auto or dry_run tier AND confidence_min ≥ MIN_CONF.
+  // min is the conservative tier-lower-bound (one weak finding drags
+  // the whole file to "simulation" or "block" tier).
+  if (r.gradient_tier === "block" || r.gradient_tier === "simulation") return false;
+  return avg >= MIN_CONF && min >= 70;
+}
+
+function passesDenyList(file: string): boolean {
+  return !DENY_PREFIXES.some((p) => file.startsWith(p) || file === p.replace(/\/$/, ""));
+}
+
+interface Patch {
+  file: string;
+  old_string: string;
+  new_string: string;
+  rationale: string;
+  confidence: number;
+}
+
+async function requestPatches(file: string, source: string, review: string): Promise<Patch[]> {
+  const prompt = `You previously produced this review of ${file}:
+
+─── REVIEW ───
+${review}
+─── END REVIEW ───
+
+The review is high-confidence and the file is eligible for auto-apply. Produce CONCRETE PATCHES as JSON so they can be applied via string replacement.
+
+RULES:
+  1. Output ONE JSON object with a "patches" array. NO prose, no markdown fences.
+  2. Each patch is {"old_string": "...", "new_string": "...", "rationale": "short", "confidence": 0-100}.
+  3. "old_string" MUST appear EXACTLY ONCE in the file (verbatim, including whitespace). If no unique anchor exists, SKIP that suggestion.
+  4. Mechanical changes only: wire a function call, add a field, remove #[allow(dead_code)], add a missing use import, rename one call-site. NO architectural rewrites. NO new modules.
+  5. Each "new_string" MUST compile in isolation with the same surrounding code. Don't introduce new dependencies.
+  6. If you cannot produce at least one high-confidence mechanical patch, output {"patches": []}.
+  7. Max 3 patches per file.
+
+─── SOURCE (${source.length} bytes) ───
+${source.slice(0, 14000)}
+─── END SOURCE ───
+
+Emit ONLY the JSON object.`;
+
+  const r = await chat({ provider: "ollama_cloud", model: MODEL, prompt, max_tokens: 2500 });
+  if (r.error || !r.content) return [];
+
+  // Strip markdown fences if model wrapped the JSON.
+  let raw = r.content.trim();
+  const fenceStart = raw.match(/^```(?:json)?\s*/);
+  if (fenceStart) raw = raw.slice(fenceStart[0].length);
+  if (raw.endsWith("```")) raw = raw.slice(0, -3).trim();
+  // Find first { and last } to extract JSON block if there's prose.
+  const first = raw.indexOf("{");
+  const last = raw.lastIndexOf("}");
+  if (first >= 0 && last > first) raw = raw.slice(first, last + 1);
+
+  try {
+    const obj = JSON.parse(raw);
+    const patches: Patch[] = (obj.patches ?? []).filter((p: any) =>
+      typeof p?.old_string === "string" &&
+      typeof p?.new_string === "string" &&
+      p.old_string !== p.new_string &&
+      p.old_string.length > 0 &&
+      typeof p?.confidence === "number"
+    ).map((p: any) => ({
+      file,
+      old_string: p.old_string,
+      new_string: p.new_string,
+      rationale: String(p.rationale ?? ""),
+      confidence: p.confidence,
+    }));
+    return patches;
+  } catch (e) {
+    log(`  ${file}: patch JSON parse failed — ${String(e).slice(0, 100)}`);
+    return [];
+  }
+}
+
+async function applyPatches(file: string, patches: Patch[]): Promise<{ applied: number; rejected: Array<{patch: Patch; reason: string}> }> {
+  const full = `${REPO}/${file}`;
+  let source = await readFile(full, "utf8");
+  const rejected: Array<{patch: Patch; reason: string}> = [];
+  let applied = 0;
+  for (const p of patches) {
+    // Confidence gate at the individual-patch level.
+    if (p.confidence < MIN_CONF) { rejected.push({patch: p, reason: `confidence ${p.confidence} < ${MIN_CONF}`}); continue; }
+    // Uniqueness gate.
+    const occurrences = source.split(p.old_string).length - 1;
+    if (occurrences === 0) { rejected.push({patch: p, reason: "old_string not found"}); continue; }
+    if (occurrences > 1) { rejected.push({patch: p, reason: `old_string appears ${occurrences}× (not unique)`}); continue; }
+    // Size gate — no patch touches > 20 lines (diff discipline).
+    const oldLines = p.old_string.split("\n").length;
+    const newLines = p.new_string.split("\n").length;
+    if (Math.max(oldLines, newLines) > 20) { rejected.push({patch: p, reason: `patch too large (${Math.max(oldLines,newLines)} lines)`}); continue; }
+    source = source.replace(p.old_string, p.new_string);
+    applied++;
+  }
+  if (applied > 0) await writeFile(full, source);
+  return { applied, rejected };
+}
+
+async function cargoCheck(): Promise<boolean> {
+  const r = await sh(["cargo", "check", "--workspace"]);
+  return r.code === 0;
+}
+
+async function gitCommit(file: string, patches: Patch[]): Promise<boolean> {
+  if (!COMMIT) { log(`  (dry-run) would commit ${file}`); return true; }
+  const addR = await sh(["git", "add", file]);
+  if (addR.code !== 0) { log(`  git add failed: ${addR.stderr.slice(0, 200)}`); return false; }
+  const msg = `auto-apply: ${patches.length} high-confidence fix${patches.length === 1 ? "" : "es"} in ${file}\n\n${patches.map(p => `- ${p.rationale} (conf ${p.confidence}%)`).join("\n")}\n\n🤖 scrum_applier.ts`;
+  const commitR = await sh(["git", "commit", "-m", msg]);
+  if (commitR.code !== 0) { log(`  git commit failed: ${commitR.stderr.slice(0, 200)}`); return false; }
+  log(`  ✓ committed ${file}`);
+  return true;
+}
+
+async function revertFile(file: string): Promise<void> {
+  await sh(["git", "checkout", "--", file]);
+}
+
+async function main() {
+  log(`starting · min_conf=${MIN_CONF} max_files=${MAX_FILES} model=${MODEL} commit=${COMMIT}`);
+
+  if (COMMIT) {
+    const headR = await sh(["git", "rev-parse", "--abbrev-ref", "HEAD"]);
+    const currentBranch = headR.stdout.trim();
+    if (currentBranch === "main") {
+      log(`refusing to run on main — create a branch first or set LH_APPLIER_BRANCH`);
+      const coR = await sh(["git", "checkout", "-b", BRANCH]);
+      if (coR.code !== 0) { log(`could not create branch ${BRANCH}: ${coR.stderr.slice(0, 200)}`); process.exit(1); }
+      log(`working branch: ${BRANCH}`);
+    } else {
+      log(`working branch: ${currentBranch}`);
+    }
+  }
+
+  const reviews = await loadLatestReviews();
+  log(`loaded ${reviews.size} latest reviews`);
+
+  const eligible = [...reviews.values()].filter(r =>
+    passesConfidenceGate(r) && passesDenyList(r.file)
+  ).sort((a, b) => (b.confidence_avg ?? 0) - (a.confidence_avg ?? 0));
+
+  log(`${eligible.length} pass confidence gate + deny-list`);
+  log(`taking top ${Math.min(MAX_FILES, eligible.length)} by confidence`);
+
+  let committedFiles = 0;
+  let revertedFiles = 0;
+
+  for (const r of eligible.slice(0, MAX_FILES)) {
+    log(`${r.file} (conf_avg=${r.confidence_avg} tier=${r.gradient_tier})`);
+    const full = `${REPO}/${r.file}`;
+    if (!existsSync(full)) { log(`  skip — file not found on disk`); continue; }
+
+    const source = await readFile(full, "utf8");
+    const patches = await requestPatches(r.file, source, r.suggestions_preview ?? "");
+
+    if (patches.length === 0) {
+      log(`  no patches produced`);
+      await auditLog({ action: "no_patches", file: r.file, reviewer_model: r.accepted_model });
+      continue;
+    }
+
+    log(`  ${patches.length} candidate patches`);
+    const { applied, rejected } = await applyPatches(r.file, patches);
+    log(`  applied ${applied}, rejected ${rejected.length}`);
+    for (const rj of rejected) log(`    ✗ ${rj.reason}`);
+
+    if (applied === 0) {
+      await auditLog({ action: "all_rejected", file: r.file, rejected: rejected.map(x => x.reason) });
+      continue;
+    }
+
+    log(`  running cargo check...`);
+    const green = await cargoCheck();
+    if (!green) {
+      log(`  ✗ build red — reverting ${r.file}`);
+      await revertFile(r.file);
+      revertedFiles++;
+      await auditLog({ action: "build_red_reverted", file: r.file, patches_applied: applied });
+      continue;
+    }
+
+    log(`  ✓ build green`);
+    const ok = await gitCommit(r.file, patches.slice(0, applied));
+    if (ok) {
+      committedFiles++;
+      await auditLog({
+        action: COMMIT ? "committed" : "dry_run_committed",
+        file: r.file,
+        patches_applied: applied,
+        patches_rejected: rejected.length,
+        confidence_avg: r.confidence_avg,
+        gradient_tier: r.gradient_tier,
+        reviewer_model: r.accepted_model,
+      });
+    }
+  }
+
+  log(`DONE · committed=${committedFiles} reverted=${revertedFiles}`);
+}
+
+await main();

tests/real-world/scrum_master_pipeline.ts

@@ -26,7 +26,7 @@ const SIDECAR = "http://localhost:3200";
 const CHUNK_SIZE = 800;
 const CHUNK_OVERLAP = 120;
 const TOP_K_CONTEXT = 5;
-const MAX_ATTEMPTS = 6;
+const MAX_ATTEMPTS = 9;
 // Files larger than this get tree-split instead of truncated. Fixes the
 // 6KB false-positive class (model claiming a field is "missing" when
 // it exists past the context cutoff).
@@ -87,21 +87,27 @@ const TARGET_FILES: string[] = process.env.LH_SCRUM_FILES
 // Local fallbacks kept for cloud-down scenarios.
 // Hot-path pipelines (scenario.ts / execution_loop) stay local per
 // Phase 20 t1_hot — this scrum is not hot path.
-const LADDER: Array<{ provider: "ollama" | "ollama_cloud"; model: string; note: string }> = [
-  { provider: "ollama_cloud", model: "kimi-k2:1t",           note: "cloud 1T — biggest available, 1.4s probe" },
-  { provider: "ollama_cloud", model: "deepseek-v3.1:671b",   note: "cloud 671B — fast reasoning (1.0s probe)" },
-  { provider: "ollama_cloud", model: "mistral-large-3:675b", note: "cloud 675B — deep analysis (0.9s probe)" },
-  { provider: "ollama_cloud", model: "gpt-oss:120b",         note: "cloud 120B — reliable workhorse (iter1 baseline)" },
-  { provider: "ollama_cloud", model: "devstral-2:123b",      note: "cloud 123B — coding specialist" },
-  // qwen3.5:397b is the deep final thinker — J's note 2026-04-24:
-  // "qwen3.5 is really smart maybe the last call use that one".
-  // When every other cloud model has produced thin output, this dense
-  // 397B reviewer is the one that tends to push through. Keeping it
-  // LAST in cloud tier, before the local fallback.
-  { provider: "ollama_cloud", model: "qwen3.5:397b",         note: "cloud 397B dense — last-ditch smart reviewer" },
-  { provider: "ollama",       model: "gpt-oss:20b",          note: "local 20B — cloud-down fallback" },
-  // kimi-k2.6 removed 2026-04-24: probe returned empty (not available
-  // on current tier). Keeping note for when pro tier upgrade lands.
+const LADDER: Array<{ provider: "ollama" | "ollama_cloud" | "openrouter"; model: string; note: string }> = [
+  { provider: "ollama_cloud", model: "kimi-k2:1t",                           note: "cloud 1T — biggest available, 1.4s probe" },
+  { provider: "ollama_cloud", model: "qwen3-coder:480b",                     note: "cloud 480B — coding specialist, 0.9s probe" },
+  { provider: "ollama_cloud", model: "deepseek-v3.1:671b",                   note: "cloud 671B — fast reasoning (1.0s probe)" },
+  { provider: "ollama_cloud", model: "mistral-large-3:675b",                 note: "cloud 675B — deep analysis (0.9s probe)" },
+  { provider: "ollama_cloud", model: "gpt-oss:120b",                         note: "cloud 120B — reliable workhorse (iter1 baseline)" },
+  { provider: "ollama_cloud", model: "qwen3.5:397b",                         note: "cloud 397B dense — deep final thinker (J 2026-04-24)" },
+  // Free-tier rescue — different provider backbone, different quota.
+  // Added 2026-04-24 after iter 5 hit repeated Ollama Cloud 502s on
+  // kimi-k2:1t. These have lower parameter counts than the Ollama
+  // Cloud rungs but high availability: if upstream is down, we still
+  // land a review instead of giving up.
+  { provider: "openrouter",   model: "openai/gpt-oss-120b:free",             note: "OpenRouter free 120B — substantive rescue, 2.8s probe" },
+  { provider: "openrouter",   model: "google/gemma-3-27b-it:free",           note: "OpenRouter free 27B — fastest rescue, 1.4s probe" },
+  { provider: "ollama",       model: "qwen3.5:latest",                       note: "local qwen3.5 — best local model per J (2026-04-24), last-resort if all cloud down" },
+  // Dropped from the ladder after 2026-04-24 probe:
+  //   - kimi-k2.6 — not available on current tier (empty response)
+  //   - devstral-2:123b — displaced by qwen3-coder:480b (better coding specialist)
+  //   - minimax-m2.7 — 400 thinking tokens, 0 content output
+  //   - openrouter qwen3-coder:free / llama-3.3 / hermes-3 — provider errors
+  //   - openrouter minimax-m2.5:free — 45s timeout
 ];
 
 type Chunk = { id: string; text: string; embedding: number[]; origin: string; offset: number };
@@ -203,10 +209,18 @@ function retrieveTopK(query_emb: number[], pool: Chunk[], k: number): Chunk[] {
     .map(x => ({ ...x.c, _score: x.score } as any));
 }
 
-// Tree-split a large file: shard it, summarize each shard against
-// the review question, merge into a scratchpad. Uses cloud because
-// the summarization step needs quality > speed. Returns the
-// scratchpad (full-file distillation) and the cloud-call count.
+// Tree-split a large file: shard it, summarize each shard into a
+// running scratchpad, THEN run a reduce step that collapses the
+// scratchpad into one file-level synthesis with shard boundaries
+// stripped. Returns the synthesis (not the raw scratchpad) so the
+// final reviewer never sees "--- shard N ---" markers and can't
+// leak them into its review title.
+//
+// Phase 21 design (aibridge/src/tree_split.rs) with the map → reduce
+// shape. Earlier version concatenated per-shard digests directly into
+// the reviewer prompt, which led to kimi-k2:1t writing review titles
+// like "Forensic Audit Report – file.rs (shard 3)" because the shard
+// markers bled through. Fix 2026-04-24 adds the reduce step.
 async function treeSplitFile(
   filePath: string,
   content: string,
@@ -216,20 +230,25 @@ async function treeSplitFile(
     const end = Math.min(i + FILE_SHARD_SIZE, content.length);
     shards.push({ from: i, to: end, text: content.slice(i, end) });
   }
-  let scratchpad = "";
+
+  // MAP — each shard produces a digest that feeds the next shard's
+  // context. Internal markers are kept to help the reducer align
+  // overlapping observations across shards; they're stripped before
+  // the reviewer sees anything.
+  let workingScratchpad = "";
   let cloud_calls = 0;
   log(`  tree-split: ${content.length} chars → ${shards.length} shards of ${FILE_SHARD_SIZE}`);
   for (const [si, shard] of shards.entries()) {
-    const prompt = `You are summarizing ONE SHARD of a source file as part of a multi-shard review. File: ${filePath}. Shard ${si + 1}/${shards.length} (bytes ${shard.from}..${shard.to}).
+    const prompt = `You are writing a SECTION of a full-file summary. File: ${filePath}. This is one piece (bytes ${shard.from}..${shard.to}) of a larger source file you are NOT seeing in its entirety right now.
 
-─────── shard source ───────
+─────── source ───────
 ${shard.text}
-─────── end shard ───────
+─────── end source ───────
 
-Scratchpad of prior shards (if empty, this is shard 1):
-${scratchpad || "(empty)"}
+Prior-piece notes so far (if empty, this is the first piece):
+${workingScratchpad || "(empty)"}
 
-Extract ONLY facts useful for reviewing this file against its PRD: function names + purposes, struct fields + types, invariants, edge cases, TODO markers, error-handling style. Under 150 words. No prose outside the extracted facts.`;
+Extract facts about the code in this piece that will help review the FULL file later: function + struct names with brief purpose, struct fields + types, invariants, TODOs, error-handling style, obvious gaps. Under 150 words. Flat facts only, no headings, no phrases like "this shard" or "in my section".`;
     const r = await chat({
       provider: "ollama_cloud",
       model: "gpt-oss:120b",
@@ -238,10 +257,47 @@ Extract ONLY facts useful for reviewing this file against its PRD: function name
     });
     cloud_calls += 1;
     if (r.content) {
-      scratchpad += `\n--- shard ${si + 1} (bytes ${shard.from}..${shard.to}) ---\n${r.content.trim()}`;
+      // Keep internal alignment markers for the reducer; stripped later.
+      workingScratchpad += `\n§${si + 1}§\n${r.content.trim()}`;
     }
   }
-  return { scratchpad, shards: shards.length, cloud_calls };
+
+  // REDUCE — collapse the per-shard digests into one coherent
+  // file-level summary. The reducer sees all digests at once and
+  // produces a single narrative the reviewer can treat as "the file".
+  // Shard markers are NOT in the output. This is what fixes the
+  // shard-leakage bug that affected both the scrum and the auditor.
+  const reducePrompt = `You are producing a SINGLE coherent summary of a Rust/TypeScript source file from a set of prior-piece notes. The notes were taken while walking the file in order but should be merged into one description of the whole file.
+
+FILE: ${filePath} (${content.length} bytes, ${shards.length} pieces)
+
+PRIOR-PIECE NOTES (markers §N§ delimit pieces but are artifacts — do not mention them):
+${workingScratchpad}
+
+Produce ONE coherent file-level summary:
+  1. One-sentence purpose of the file.
+  2. Key public types / functions / constants (names + one-line purpose each).
+  3. Known gaps, TODOs, or error-handling inconsistencies the notes surfaced.
+  4. Obvious invariants the file relies on.
+
+Do NOT say "piece 1" or "shard N" or "section" — present the summary as if you read the whole file at once. Under 600 words.`;
+
+  const reduced = await chat({
+    provider: "ollama_cloud",
+    model: "gpt-oss:120b",
+    prompt: reducePrompt,
+    max_tokens: 900,
+  });
+  cloud_calls += 1;
+  const synthesis = reduced.content?.trim() ?? "";
+
+  // Safety: if the reducer returned thin output, fall back to the
+  // raw scratchpad stripped of markers — better than nothing.
+  const final = synthesis.length > 200
+    ? synthesis
+    : workingScratchpad.replace(/§\d+§\n/g, "").trim();
+
+  return { scratchpad: final, shards: shards.length, cloud_calls };
 }
 
 async function reviewFile(