refactor(platform): NodeKit standard tree phase 2 — full physical migration

.claude/rules/backend_contract_migration.md

@@ -1,9 +1,9 @@
 ---
 paths:
-  - "convex/**/*.ts"
-  - "server/**/*.ts"
+  - "backend/convex/**/*.ts"
+  - "workers/node/**/*.ts"
   - "public/proto/*.html"
-  - "src/**/*.{ts,tsx}"
+  - "apps/web/src/**/*.{ts,tsx}"
 related_: [live_dom_verification, agentic_reliability, owner_mode_end_to_end, completion_traceability]
 ---
 
@@ -41,7 +41,7 @@ contract, ship in **three PRs**, not one:
 - Old callers keep working; new callers can opt in.
 - For a rename: export both names, both pointing at the same handler.
   ```ts
-  // convex/events.ts
+  // backend/convex/events.ts
   const composeAnswerHandler = mutation({ /* ... */ });
   export const composeAnswer = composeAnswerHandler;
   export const askAgent = composeAnswerHandler;  // legacy alias
@@ -101,9 +101,9 @@ If verification fails at any step, do not proceed to the next step.
 
 ## Contract path discipline (added 2026-05-27)
 
-**Every Convex contract MUST specify the deployed path as `<filename>:<exportName>`, not just the export name.** Convex resolves paths by filename — putting `requestSignInLink` in `convex/users.ts` deploys it at `users:requestSignInLink`, not `events:requestSignInLink`, regardless of what the contract doc says.
+**Every Convex contract MUST specify the deployed path as `<filename>:<exportName>`, not just the export name.** Convex resolves paths by filename — putting `requestSignInLink` in `backend/convex/users.ts` deploys it at `users:requestSignInLink`, not `events:requestSignInLink`, regardless of what the contract doc says.
 
-Case study — PR #407/#409 hotfix (2026-05-27): the Step 8 contract specified mutations as `events:requestSignInLink` / `events:verifySignInToken` / `events:listMyEvents` to match the existing scratchnode namespace. The implementing agent put them in `convex/users.ts` (sensible — `events.ts` was overloaded). Deployed paths became `users:*`. Frontend and dogfood script kept calling `events:*` — silent function-not-found at runtime. Convex's HTTP API masks this as a generic "Server Error" message, hiding the mismatch from CI.
+Case study — PR #407/#409 hotfix (2026-05-27): the Step 8 contract specified mutations as `events:requestSignInLink` / `events:verifySignInToken` / `events:listMyEvents` to match the existing scratchnode namespace. The implementing agent put them in `backend/convex/users.ts` (sensible — `events.ts` was overloaded). Deployed paths became `users:*`. Frontend and dogfood script kept calling `events:*` — silent function-not-found at runtime. Convex's HTTP API masks this as a generic "Server Error" message, hiding the mismatch from CI.
 
 ### Required contract format
 
@@ -113,13 +113,13 @@ events:requestSignInLink({ email }) → { ok }
 
 // ✅ GOOD — deployed path is explicit
 users:requestSignInLink({ email }) → { ok }
-// implemented in convex/users.ts as `export const requestSignInLink = mutation({...})`
+// implemented in backend/convex/users.ts as `export const requestSignInLink = mutation({...})`
 ```
 
 If the contract author wants the path to be `events:*` despite the implementation living in another file, the implementing PR must add an explicit re-export:
 
 ```ts
-// convex/events.ts
+// backend/convex/events.ts
 export { requestSignInLink, verifySignInToken, listMyEvents } from "./users";
 ```
 

.claude/rules/dogfood_verification.md

@@ -88,9 +88,9 @@ Runs a full Scribe-like dogfood session locally and makes it UI-verifiable at `/
 
 | Layer | Changes to | Deployed via | Visible when |
 |-------|-----------|-------------|--------------|
-| React components | `.tsx` in `src/` | `vite build` | Preview server restarted |
-| Convex functions | `.ts` in `convex/` | `npx convex deploy` | Backend redeployed |
-| Convex schema | `convex/schema.ts` | `npx convex deploy` | Backend redeployed |
+| React components | `.tsx` in `apps/web/src/` | `vite build` | Preview server restarted |
+| Convex functions | `.ts` in `backend/convex/` | `npx convex deploy` | Backend redeployed |
+| Convex schema | `backend/convex/schema.ts` | `npx convex deploy` | Backend redeployed |
 | MCP tools | `packages/mcp-local/` | `npx tsc` in package | MCP server restarted |
 | Stored data | DB records | Data migration or mutation | After migration runs |
 | Dashboard HTML | `briefHtml.ts` / `html.ts` | Inline — restart server | Server restarted |

.claude/rules/eval_flywheel.md

@@ -3,7 +3,7 @@
 Continuous self-improving eval loop for the NodeBench search pipeline. The system judges itself, diagnoses failures, fixes them, and loops until 100%.
 
 ## When to activate
-- After any change to `server/routes/search.ts` or search-related tools
+- After any change to `workers/node/routes/search.ts` or search-related tools
 - User says "run eval", "flywheel", "judge loop", "reach 100%"
 - After completing any implementation sprint touching search, entity enrichment, or result rendering
 - Automatically after deploying search changes
@@ -109,10 +109,10 @@ Always do deep research before declaring a blocker permanent. The system should
 ## Key files
 - `packages/mcp-local/src/benchmarks/searchQualityEval.ts` — Eval harness (100+ queries, Gemini judge)
 - `packages/mcp-local/src/benchmarks/llmJudgeEval.ts` — Chained pipeline eval (tool A → tool B)
-- `server/routes/search.ts` — Search route with 4-layer grounding pipeline
+- `workers/node/routes/search.ts` — Search route with 4-layer grounding pipeline
 - `packages/mcp-local/src/tools/entityEnrichmentTools.ts` — Entity enrichment MCP tools
 - `packages/mcp-local/src/tools/webTools.ts` — web_search implementation
-- `convex/tools/media/linkupSearch.ts` — Linkup search (Convex-side)
+- `backend/convex/tools/media/linkupSearch.ts` — Linkup search (Convex-side)
 
 ## Anti-patterns
 - Declaring done at 80% because "the remaining 20% is hard"

.claude/rules/forecasting_os.md

@@ -1,12 +1,12 @@
 ---
 paths:
-  - "convex/domains/forecasting/**"
-  - "convex/workflows/dailyLinkedInPost.ts"
-  - "src/features/research/components/Forecast*"
-  - "src/features/research/components/CalibrationPlot*"
-  - "src/features/research/components/BrierTrendChart*"
-  - "src/features/research/components/EvidenceTimeline*"
-  - "src/features/research/components/TraceBreadcrumb*"
+  - "backend/convex/domains/forecasting/**"
+  - "backend/convex/workflows/dailyLinkedInPost.ts"
+  - "apps/web/src/features/research/components/Forecast*"
+  - "apps/web/src/features/research/components/CalibrationPlot*"
+  - "apps/web/src/features/research/components/BrierTrendChart*"
+  - "apps/web/src/features/research/components/EvidenceTimeline*"
+  - "apps/web/src/features/research/components/TraceBreadcrumb*"
   - "packages/mcp-local/src/tools/forecastingTools.ts"
 related_: [reexamine_process, analyst_diagnostic, completion_traceability, reexamine_resilience]
 ---
@@ -32,13 +32,13 @@ related_: [reexamine_process, analyst_diagnostic, completion_traceability, reexa
 ## Key files
 | File | Purpose |
 |------|---------|
-| `convex/domains/forecasting/forecastManager.ts` | CRUD + 6 public dashboard queries |
-| `convex/domains/forecasting/signalMatcher.ts` | Deterministic signal↔forecast cross-reference |
-| `convex/domains/forecasting/traceWrapper.ts` | TRACE-wrapped forecast refresh (6 audit steps) |
-| `convex/domains/forecasting/scoringEngine.ts` | Brier + log scoring, proper scoring rules |
-| `convex/domains/forecasting/schema.ts` | 5 tables: forecasts, forecastEvidence, forecastResolutions, forecastUpdateHistory, forecastCalibrationLog |
-| `convex/workflows/dailyLinkedInPost.ts` | LinkedIn pipeline with Δ badges, evidence links, TRACE |
-| `src/features/research/components/ForecastCockpit.tsx` | Dashboard assembler (CalibrationPlot, BrierTrendChart, ForecastCard) |
+| `backend/convex/domains/forecasting/forecastManager.ts` | CRUD + 6 public dashboard queries |
+| `backend/convex/domains/forecasting/signalMatcher.ts` | Deterministic signal↔forecast cross-reference |
+| `backend/convex/domains/forecasting/traceWrapper.ts` | TRACE-wrapped forecast refresh (6 audit steps) |
+| `backend/convex/domains/forecasting/scoringEngine.ts` | Brier + log scoring, proper scoring rules |
+| `backend/convex/domains/forecasting/schema.ts` | 5 tables: forecasts, forecastEvidence, forecastResolutions, forecastUpdateHistory, forecastCalibrationLog |
+| `backend/convex/workflows/dailyLinkedInPost.ts` | LinkedIn pipeline with Δ badges, evidence links, TRACE |
+| `apps/web/src/features/research/components/ForecastCockpit.tsx` | Dashboard assembler (CalibrationPlot, BrierTrendChart, ForecastCard) |
 | `packages/mcp-local/src/tools/forecastingTools.ts` | 9 MCP tools |
 
 ## Conventions

.claude/rules/gemini_qa_loop.md

@@ -20,7 +20,7 @@ npx vite build
 npx vite preview --host 127.0.0.1 --port 4173 &
 
 # 3. Capture screenshots via e2e test
-BASE_URL=http://127.0.0.1:4173 npx playwright test tests/e2e/full-ui-dogfood.spec.ts --project=chromium --workers=1
+BASE_URL=http://127.0.0.1:4173 npx playwright test evals/e2e/full-ui-dogfood.spec.ts --project=chromium --workers=1
 
 # 4. Publish screenshots to public/dogfood/
 npm run dogfood:publish
@@ -92,9 +92,9 @@ Pattern: **Pro → Flash → Flash → Flash → Pro → Flash → Flash → Fla
 - 60% cost reduction while maintaining Pro-quality baseline analysis
 
 ### Config files
-- `convex/domains/dogfood/screenshotQa.ts` — Screenshot QA action
-- `convex/domains/dogfood/videoQa.ts` — Video QA action
-- `convex/domains/dogfood/videoQaQueries.ts` — `getLatestProAnalysis` query for reference injection
+- `backend/convex/domains/dogfood/screenshotQa.ts` — Screenshot QA action
+- `backend/convex/domains/dogfood/videoQa.ts` — Video QA action
+- `backend/convex/domains/dogfood/videoQaQueries.ts` — `getLatestProAnalysis` query for reference injection
 
 ## Multi-Variant Coverage
 
@@ -125,12 +125,12 @@ This returns the step-by-step workflow chain with tool references and shell comm
 
 | File | Purpose |
 |------|---------|
-| `convex/domains/dogfood/screenshotQa.ts` | Screenshot QA action (Gemini Flash + Jony Ive prompts) |
-| `convex/domains/dogfood/videoQa.ts` | Video QA action (Gemini Flash + Jony Ive prompts) |
+| `backend/convex/domains/dogfood/screenshotQa.ts` | Screenshot QA action (Gemini Flash + Jony Ive prompts) |
+| `backend/convex/domains/dogfood/videoQa.ts` | Video QA action (Gemini Flash + Jony Ive prompts) |
 | `scripts/ui/runDogfoodGeminiQa.mjs` | CLI orchestrator for QA pipeline |
 | `scripts/ui/recordDogfoodWalkthrough.mjs` | Playwright video recorder |
 | `scripts/ui/publishDogfoodGallery.mjs` | Screenshot publisher (variant-aware manifest) |
-| `tests/e2e/full-ui-dogfood.spec.ts` | E2e 4-variant screenshot capture test |
+| `evals/e2e/full-ui-dogfood.spec.ts` | E2e 4-variant screenshot capture test |
 | `public/dogfood/qa-results.json` | QA score history |
 | `.tmp/dogfood-gemini-qa/*.json` | Latest QA results (screens + video) |
 | `shared/llm/modelCatalog.ts` | Model catalog with Gemini defaults |

.claude/rules/grounded_eval.md

@@ -16,7 +16,7 @@ Before extraction, check source quality:
 - **medium**: 1-2 snippets — extract conservatively, flag as limited
 - **low**: 0 snippets — return "insufficient data" template, do NOT generate
 
-Implementation: `retrievalConfidence` in `server/routes/search.ts`
+Implementation: `retrievalConfidence` in `workers/node/routes/search.ts`
 
 ### Layer 2: Claim-Level Grounding Filter
 After Gemini extraction, verify each claim against source text:
@@ -57,7 +57,7 @@ Implementation: `sourceIdx` field on signals, changes, risks in response
 - If judge variance exceeds 10% across runs, use majority vote (3x calls)
 
 ## Key files
-- `server/routes/search.ts` — Layers 1-2 and 4 (retrieval, filter, citations)
+- `workers/node/routes/search.ts` — Layers 1-2 and 4 (retrieval, filter, citations)
 - `packages/mcp-local/src/benchmarks/searchQualityEval.ts` — Layer 3 (grounded judge)
 - `packages/mcp-local/src/benchmarks/llmJudgeEval.ts` — Chain coherence criterion
 

.claude/rules/live_dom_verification.md

@@ -61,7 +61,7 @@ hydrates + Convex queries resolve. That means:
   return 200 not 404. Catches landmines (a) and (c).
   Does NOT prove: routes actually render the right component.
 
-- **Tier B (tests/e2e/live-smoke.spec.ts via `npm run live-smoke`)** —
+- **Tier B (evals/e2e/live-smoke.spec.ts via `npm run live-smoke`)** —
   Playwright loads each URL in a real browser, waits for hydration,
   asserts DOM nodes exist (e.g. "Link not found" on `/share/dummy`,
   `<h1>` on landing, recovery CTAs visible).

.claude/rules/pipeline_operational_standard.md

@@ -34,7 +34,7 @@ instrument → judge → persist → surface → measure → regress.
 `reportsTokenCounts`, `capturedSources`, `emitStatusIsTerminal`.
 
 Order is stable — dashboards rely on it. Add a gate by extending `GATE_ORDER`
-in `server/pipeline/diligenceJudge.ts`.
+in `workers/node/pipeline/diligenceJudge.ts`.
 
 ## Verdict tiers (bounded enum)
 - `verified` — 0 failures, ≤ 2 skipped
@@ -45,7 +45,7 @@ in `server/pipeline/diligenceJudge.ts`.
 ## Verification floor
 1. `npx convex codegen`
 2. `npx tsc --noEmit`
-3. `npx vitest run server/pipeline/diligenceJudge.test.ts server/pipeline/diligenceProjectionWriter.test.ts`
+3. `npx vitest run workers/node/pipeline/diligenceJudge.test.ts workers/node/pipeline/diligenceProjectionWriter.test.ts`
 4. `npm run build`
 5. `npm run dogfood:verify:smoke` when the UI changed
 

.claude/rules/pre_release_review.md

@@ -106,7 +106,7 @@ If Convex is NOT deployed:
 - [ ] "Sign in" messages appear where expected
 
 ### Layer 10: WebSocket Gateway (< 3 min)
-If gateway is running (`npx tsx server/index.ts`):
+If gateway is running (`npx tsx workers/node/index.ts`):
 - [ ] `GET /health` returns 200
 - [ ] `GET /mcp/health` returns session count
 - [ ] WebSocket connects with valid API key
@@ -133,8 +133,8 @@ Test in at least one non-Chrome browser (Safari, Firefox, or Edge):
 
 ### Layer 13: Regression Risks (< 2 min)
 Check these known fragile areas:
-- [ ] Voice server: if `server/index.ts` was modified, verify voice WebSocket still works
-- [ ] Schema changes: if `convex/schema.ts` was modified, verify migration compatibility
+- [ ] Voice server: if `workers/node/index.ts` was modified, verify voice WebSocket still works
+- [ ] Schema changes: if `backend/convex/schema.ts` was modified, verify migration compatibility
 - [ ] Tool count: grep for hardcoded "289", "297", "304" — must all match current reality
 - [ ] Print: if Decision Memo or Postmortem was modified, verify print stylesheet (if exists)
 - [ ] OG tags: if `index.html` was modified, verify meta tags are correct

.claude/rules/product_design_dogfood.md

@@ -1,8 +1,8 @@
 ---
 paths:
-  - "src/**/*.tsx"
-  - "src/**/*.css"
-  - "tests/e2e/**/*.ts"
+  - "apps/web/src/**/*.tsx"
+  - "apps/web/src/**/*.css"
+  - "evals/e2e/**/*.ts"
   - "scripts/ui/**/*.mjs"
 related_: [analyst_diagnostic, dogfood_verification, reexamine_design_reduction, completion_traceability]
 ---

.claude/rules/reference_attribution.md

@@ -58,7 +58,7 @@ from topic files rather than persisting it directly." Honest > embellished.
 ## What this rule enforces
 
 - New architecture docs have a "Prior art" section or the PR is rejected
-- New server/pipeline/ or convex/domains/ modules have a file header comment linking their doc
+- New workers/node/pipeline/ or backend/convex/domains/ modules have a file header comment linking their doc
 - When copying a Claude Code skill, cite `.claude/skills/<origin>` in the new skill's frontmatter
 
 ## Anti-patterns

.claude/rules/self_improvement_loop.md

@@ -1,13 +1,13 @@
 # Self-Improvement Loop — Operating Manual
 
 The agent brain for NodeBench's continuous improvement flywheel. Read this on every loop cycle
-(manual or scheduled). The deterministic substrate is `scripts/improvement-loop/`; this rule is
+(manual or scheduled). The deterministic substrate is `adw/improvement-loop/`; this rule is
 how the agent drives it. Canonical design: `docs/architecture/SELF_IMPROVEMENT_LOOP.md`.
 
 ## Operating model: bounded + goal-driven (NOT "never stop")
 
-This loop is governed by the Self-Directed Development OS in [`goals/README.md`](../../goals/README.md)
-and the hard gates in [`goals/HARD_GATES.md`](../../goals/HARD_GATES.md). The lesson is **not**
+This loop is governed by the Self-Directed Development OS in [`adw/goals/README.md`](../../goals/README.md)
+and the hard gates in [`adw/goals/HARD_GATES.md`](../../goals/HARD_GATES.md). The lesson is **not**
 "let the agent run forever" — it is: closed goal, reviewable definition of done, focused subagents,
 hard gates, batch feedback.
 
@@ -18,7 +18,7 @@ hard gates, batch feedback.
 
 - **Cadence:** daily small-loop (propose ONE bounded next step; only tiny CI-gated detector fixes
   auto-ship, ≤3/day) + weekly self-review (propose issues/cuts/goals, never auto-add features).
-  Substantive work becomes a Goal Card (`goals/<surface>/NNN-slug.md`) the founder approves.
+  Substantive work becomes a Goal Card (`adw/goals/<surface>/NNN-slug.md`) the founder approves.
 - **Operating rule:** Human sets the *why* + boundary; agent explores the *how*; tests decide; docs preserve.
 - **Hard gates:** prod deploy, destructive migrations, auth, billing, public/private permission
   rules, data deletion, legal/privacy copy, wiki publish, host/mod privileges, secrets → propose only.
@@ -32,7 +32,7 @@ hard gates, batch feedback.
 
 ### 1. OBSERVE + SCORE
 ```bash
-node scripts/improvement-loop/run-cycle.mjs --effort-budget 3
+node adw/improvement-loop/run-cycle.mjs --effort-budget 3
 ```
 This runs `scan.mjs`, writes `backlog.latest.json`, selects the top auto-safe opportunity, and
 appends a cycle to `ledger.json`.
@@ -84,7 +84,7 @@ If the change would touch any of these, it is **human-gated** — queue it, do N
 - Shipping a fix for a false-positive scanner hit (validate first).
 - Auto-shipping a human-gated change.
 - Claiming "live" on a green build (live-DOM verify first — `live_dom_verification.md`).
-- Building a parallel loop instead of extending `scripts/improvement-loop/` + the existing dogfood/eval scripts.
+- Building a parallel loop instead of extending `adw/improvement-loop/` + the existing dogfood/eval scripts.
 
 ## Related rules
 - `flywheel_continuous` · `self_building_loop` · `eval_flywheel` · `analyst_diagnostic`

.claude/rules/telemetry_trajectory.md

@@ -10,7 +10,7 @@ Make telemetry and agent step trajectories measurable, debuggable, and beautiful
 
 ## Dev-side telemetry (internal)
 
-### Search trace (server/routes/search.ts)
+### Search trace (workers/node/routes/search.ts)
 Every search request emits a `trace` array with structured steps:
 ```typescript
 interface TraceStep {
@@ -83,13 +83,13 @@ Every search result should be traceable back to:
 5. What the user should verify independently
 
 ## Key files
-- `src/features/controlPlane/components/SearchTrace.tsx` — User-facing trace UI
-- `src/features/agents/components/FastAgentPanel/StepTimeline.tsx` — Agent step timeline
-- `src/features/agents/components/FastAgentPanel/FastAgentPanel.ParallelTaskTimeline.tsx` — Parallel execution
-- `src/features/monitoring/views/AgentTelemetryDashboard.tsx` — Telemetry dashboard
-- `src/features/trajectory/types.ts` — Trajectory score types
-- `convex/domains/agents/traceTypes.ts` — TRACE framework types
-- `server/routes/search.ts` — Search trace emission
+- `apps/web/src/features/controlPlane/components/SearchTrace.tsx` — User-facing trace UI
+- `apps/web/src/features/agents/components/FastAgentPanel/StepTimeline.tsx` — Agent step timeline
+- `apps/web/src/features/agents/components/FastAgentPanel/FastAgentPanel.ParallelTaskTimeline.tsx` — Parallel execution
+- `apps/web/src/features/monitoring/views/AgentTelemetryDashboard.tsx` — Telemetry dashboard
+- `apps/web/src/features/trajectory/types.ts` — Trajectory score types
+- `backend/convex/domains/agents/traceTypes.ts` — TRACE framework types
+- `workers/node/routes/search.ts` — Search trace emission
 - `packages/mcp-local/src/benchmarks/searchQualityEval.ts` — Eval harness
 
 ## Related rules