pec: TASK-001 output

docs/program-execution/TASK-001.md

@@ -0,0 +1,32 @@
+# TASK-001 — Canonical capability resolver
+
+Campaign: `pe-router-capability-safety` · Base SHA `14c7b83` · Branch `pec/w0/task-001`
+
+## What changed
+
+- `src/types.ts` — added `ResolvedCapabilities` (searchRequired, searchPolicySource,
+  visionRequired) next to `SearchPolicyResolution`, so the resolver's output is a
+  first-class shared type.
+- `src/matrices/capabilities.ts` (new) — `VISION_TASKS` exported as the single
+  authority for vision task types (VISUAL_QA, SCREENSHOT_ANALYSIS,
+  LAYOUT_VALIDATION) and the canonical `resolveCapabilities(task)` resolver.
+  Search truth delegates to `resolveSearchPolicy` — exactly one implementation of
+  the search rule; `searchPolicySource` preserves the existing EXPLICIT /
+  TASK_DEFAULT enum shape.
+- `src/index.ts` — `resolveRoute` now consumes `resolveCapabilities` instead of
+  re-deriving the search policy locally; the private `VISION_TASKS` constant was
+  removed in favor of the exported one. The fail-closed search+vision guard and
+  the dispatch behavior are unchanged.
+- `tests/capabilities.test.ts` (new) — focused coverage: explicit true/false
+  flags report EXPLICIT source, undefined flags report TASK_DEFAULT with the
+  TaskType default, exactly the vision types are vision-required, and a vision
+  task with an explicit search=false keeps both truths distinct.
+
+## Validation (run on the finished tree)
+
+- `npm run verify:types` — PASS (tsc --noEmit)
+- `npm test` — PASS (23 files, 165 tests)
+- `npm run lint` — PASS (eslint src/)
+
+Existing search-policy tests and routing behavior are unchanged: the resolver
+reuses `resolveSearchPolicy` and the same guard fires in the same cases.

docs/program-execution/TASK-002.md

@@ -0,0 +1,36 @@
+# TASK-002 — Fail unsupported capability combinations
+
+Campaign: `pe-router-capability-safety-v4` · Stacked on TASK-001
+
+## What changed
+
+- `src/matrices/capabilities.ts` — added `VisionInputRequiredError`
+  (code `VISION_INPUT_REQUIRED`) and `assertSupportedCapabilities(task,
+  capabilities)`, which refuses before any reservation, circuit permit, or
+  provider dispatch:
+  - visionRequired + searchRequired → `UNSUPPORTED_CAPABILITY_COMBINATION`
+    (no provider serves search and vision together);
+  - images on a non-vision task type → `UNSUPPORTED_CAPABILITY_COMBINATION`
+    (images are only consumed by the vision branch);
+  - a vision task with no images → `VISION_INPUT_REQUIRED`.
+- `src/index.ts` — `resolveRoute` now calls `assertSupportedCapabilities`
+  instead of the narrower images-present guard; `dispatchProvider` gained the
+  vision dispatch invariant (a vision task must dispatch on the OpenRouter
+  vision plane, never Perplexity or the general path).
+- `tests/routing-matrix.test.ts` — matrix rows added: screenshot analysis
+  without images fails closed, content generation with images fails closed,
+  visual QA with images routes to vision; FAIL_CLOSED assertions now check the
+  failure code (`UNSUPPORTED_CAPABILITY_COMBINATION` or
+  `VISION_INPUT_REQUIRED`).
+- `tests/search-policy-dispatch.test.ts` — updated the two legacy cases the
+  contract supersedes: a visual task with no images is now refused
+  (`VISION_INPUT_REQUIRED`) instead of routing with a defaulted image count,
+  and a visual task with no images plus explicit search is now an
+  `UNSUPPORTED_CAPABILITY_COMBINATION` instead of a plain search request; the
+  image-count regression guard now asserts the no-image refusal explicitly.
+
+## Validation (run on the finished tree)
+
+- `npm run verify:types` — PASS
+- `npm test` — PASS (23 files, 169 tests)
+- `npm run lint` — PASS

docs/program-execution/TASK-003.md

@@ -0,0 +1,25 @@
+# TASK-003 — Capability truth in RoutingDecision
+
+Campaign: `pe-router-capability-safety-v4` · Stacked on TASK-002
+
+## What changed
+
+- `src/types.ts` — `RoutingResolution` gained `visionRequired: boolean`
+  (documented), inherited by `RoutingDecision`, alongside the existing
+  `searchRequired` and `searchPolicySource` audit fields from the merged
+  search-policy work. The `SearchPolicySource` enum shape is unchanged.
+- `src/index.ts` — `resolveRoute` populates `visionRequired` from the
+  canonical `resolveCapabilities` output, so every decision reports the full
+  capability truth without re-deriving it at the call site.
+- `tests/routing-matrix.test.ts` — added a regression test asserting
+  capability truth on every routing decision: a vision route reports
+  `visionRequired: true` with `searchRequired: false` and EXPLICIT source
+  (proving the SEO_CONTENT_BLUEPRINT-style audit shape without inferring
+  policy from the model name), a general route reports `visionRequired:
+  false`, and a search route reports `visionRequired: false` on Perplexity.
+
+## Validation (run on the finished tree)
+
+- `npm run verify:types` — PASS
+- `npm test` — PASS
+- `npm run lint` — PASS

docs/program-execution/TASK-004.md

@@ -0,0 +1,22 @@
+# TASK-004 — Required routing matrix regression tests
+
+Campaign: `pe-router-capability-safety-v4` · Stacked on TASK-003
+
+## What changed
+
+- `tests/routing-matrix.test.ts` — completed the required routing matrix
+  from the Website Contract: added the VISUAL_QA with **multiple** images
+  case (row U) routing to vision with requiresSearch=false. Together with
+  the rows landed in TASK-002, the matrix now covers every required row:
+  STRATEGIC_REASONING with false → general and true → search;
+  COMPETITOR_RESEARCH undefined → search and false → general;
+  SCREENSHOT_ANALYSIS with image+false → vision, image+true → unsupported
+  combination, no images → vision input required; CONTENT_GENERATION with
+  images → unsupported combination; VISUAL_QA with multiple images+false →
+  vision. All prior matrix rows stay green.
+
+## Validation (run on the finished tree)
+
+- `npm run verify:types` — PASS
+- `npm test` — PASS
+- `npm run lint` — PASS

docs/program-execution/TASK-005.md

@@ -0,0 +1,28 @@
+# TASK-005 — Bump version to the 1.3.0 release line
+
+Campaign: `pe-router-capability-safety-v4` · Stacked on TASK-004
+
+## What changed
+
+- `package.json` and `package-lock.json` (root + packages entry) bumped from
+  1.1.3 to **1.3.0**.
+
+Operator decision (2026-08-18, mid-campaign): the release line is 1.3, not
+1.1 — the search-policy audit work plus this campaign's capability-safety
+work ship as 1.3.0, tagged by the operator after this campaign's delivery PR
+merges. The intermediate 1.2.0 bump from PR #46 and the interim 1.1.3 pin
+(PR #56) are superseded by this single forward jump; a separate version-fix
+PR would have conflicted with this change on the same lines, so the campaign
+itself is the 1.3.0 vehicle. Registry-state verification happens at the
+operator's tag step (GitHub Packages authentication is operator-held).
+
+## Validation (run on the finished tree)
+
+- `npm run verify:all` — build, types, declarations, lint, boundary lint,
+  171+ tests, audit, package smoke
+- `npm run verify:package` — packed-tarball smoke install passes
+
+The packed artifact (npm pack) and the declaration-consumer fixture inside
+verify:all stand as the local release proof; full disposable Website-Bot and
+SEO-Bot consumer installs run at the operator's release step before the
+1.3.0 tag push, per the contract's release sequence.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@quantum-l9/llm-router",
-  "version": "1.1.3",
+  "version": "1.3.0",
   "type": "module",
   "description": "Reusable multi-provider LLM routing module with governed l9-graphiti-memory hydration, task-to-model routing, budgets, search, vision, and provider resilience.",
   "main": "dist/index.js",