Add bounded knowledge retrieval

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 0d8b7764-f15a-49ea-8d50-d9147334f8af

microsoft/skills/review/al-code-review.md

@@ -65,7 +65,10 @@ The worklist is the list of sub-skills judged relevant by the previous step. Eve
 
 The Action step consists of **discrete leaf invocations**, not one combined generation. Invocation scheduling belongs to the orchestrator: independent leaves may run serially or concurrently, but their evaluation contexts and findings-reports remain isolated. Concretely this means:
 
-- **Isolate leaf invocations when the host supports it.** For fast/small models, each sub-skill SHOULD run in a fresh model call or child context containing only the task input, READ/DO contracts, the leaf instructions, a domain-filtered slice of the current knowledge index, and articles that leaf worklists. Preserve each index row's exact `path`; the leaf must copy references from that slice. The coordinator then collects the resulting JSON. This is the preferred fast-model profile: it bounds context, prevents later leaves from being skipped as attention is exhausted, and removes any reason to synthesize article paths.
+- **Isolate leaf invocations when the host supports it.** Each sub-skill SHOULD run in a fresh model call or child context containing only its assigned source paths, READ/DO contracts, the leaf instructions, the complete bounded domain catalog per READ, and articles that leaf worklists. Preserve each catalog row's exact `path`; the leaf must copy references from that catalog.
+- **Keep run artifacts private.** Before dispatch, allocate a new GUID-named directory under the current session's artifact directory and a distinct scratch/report child directory for every leaf. Pass a leaf only its own assigned source paths and child directory, never the run root or sibling paths. A leaf MUST NOT discover, enumerate, read, modify, or delete sibling artifacts. Do not reuse a prior run directory, and do not clean up any run artifact until every leaf has finished and consolidation is complete.
+- **Use the exact Task return as the report.** Capture each leaf's exact return as the primary transport and apply DO's consumer acceptance gate before rollup. Worker-side persistence of the same report in its private directory is optional and redundant; a missing report file does not invalidate an otherwise valid exact return.
+- **Treat automatic output spills as host-owned.** If the host reports that a Task return was automatically spilled, the coordinator MAY read that file read-only only at the exact path returned by the tool. Never modify, delete, enumerate around, or reuse an automatic spill path. Never bypass a content-exclusion or access denial.
 - Treat each sub-skill in the worklist as its own pass: read the sub-skill's instructions, apply its Source → Relevance → Worklist → Action steps to the orchestrator-supplied inputs, and produce that sub-skill's complete findings-report independently.
 - Do not collapse multiple sub-skills into one shared reasoning step. Each sub-skill has a distinct knowledge subset and a distinct evaluation procedure; sharing one rolled-up scan dilutes per-skill attention and causes leaves to silently underreport (this has been observed in production: leaf skills returned empty `findings[]` while their standalone runs against the same diff produced multiple matches).
 - The agent self-review pass is its own final iteration. Begin it only after every sub-skill in the worklist has completed and its sub-result is recorded.
@@ -77,8 +80,8 @@ The Action step consists of **discrete leaf invocations**, not one combined gene
 For each sub-skill in the worklist:
 
 1. Invoke the sub-skill with the orchestrator's inputs, passing only the subset each sub-skill declares in its `inputs`.
-2. Capture the sub-skill's complete findings-report verbatim and append it to `sub-results`.
-3. If the sub-skill's `outcome` is `failed`, stop here for this sub-skill: its findings are not reliable per the DO contract and MUST NOT be copied into the super-skill's top-level `findings[]` or counted in `summary.counts`.
+2. Capture the exact Task return and validate it against DO's consumer acceptance gate before accepting it. Preserve an invalid raw return unchanged in the leaf's private artifacts or host log; do not reconstruct or repair it. Record a separate failed validation result with no findings for rollup.
+3. Append the accepted findings-report, or the separate failed validation result, to `sub-results`. If its `outcome` is `failed`, stop here for this sub-skill: its findings are not reliable per the DO contract and MUST NOT be copied into the super-skill's top-level `findings[]` or counted in `summary.counts`.
 4. Otherwise, compare each entry from the sub-skill's `findings[]` with findings already rolled up. Two findings are duplicates when they point to the same file and overlapping line/range and prescribe materially the same correction, even when their knowledge-file IDs differ. Merge duplicates instead of appending both: keep the more specific domain owner, preserve that finding's optional `domain` field verbatim (including its absence), use its reference as `references[0]` and therefore as `id`, append the other references as supporting references, keep the highest severity and confidence justified by either report, and preserve one self-contained message. Article and leaf ownership notes decide specificity; do not choose by execution order.
 5. Append each non-duplicate finding, setting `from-sub-skill` to the sub-skill's `skill.id` and preserving its optional `domain` field verbatim, including its absence. For non-citation findings (those whose `id` is a skill-defined slug rather than a reference path), prefix `id` with `<from-sub-skill>:` to prevent collisions across sub-skills. Other finding fields are preserved.
 
@@ -122,7 +125,10 @@ Calculate `summary.counts` from the final top-level `findings[]`, after failed s
 
 Derive `outcome` using the DO rollup rules. `outcome-reason` is populated for `partial` and `failed` and SHOULD summarize per-sub-skill state, for example: *"al-security-review failed (tool timeout); al-performance-review completed."*
 
-Before emitting the rollup, apply DO's reference-integrity gate to every nested and top-level finding. Every knowledge-backed ID/reference path must exist in the live checkout, must have been opened by the producing leaf, and must be copied verbatim rather than synthesized. Treat a sub-result containing an unverifiable citation as failed and exclude its findings from the top-level rollup.
+Before emitting the rollup, apply DO's consumer acceptance gate to every nested
+and top-level finding. Treat an invalid sub-result as failed and exclude all of
+its findings from the top-level rollup. Preserve its exact raw payload
+separately; never reconstruct it into a success-shaped report.
 
 ## Output