docs(memory): preserve two remaining stranded lessons from stale worktrees

.claude/agent-memory/orchestrator/delegate-may-lack-bash-tool-verify-its-git-claims.md

@@ -27,6 +27,27 @@ expected, not as evasion, and run them yourself the moment the delegate returns:
 - `git -C <worktree> diff --cached --numstat` after staging, to confirm `.gitattributes`
   normalisation did what you expected.
 
+**`task-researcher` also has no Bash tool, which makes `gh` work undelegatable (verified 2026-09-12,
+issue #743 preparation).** Its granted surface is Read, Grep, Glob, WebFetch and Write only. A research
+prompt that says "read issues 592, 511 and 571 including their closing comments" is therefore
+unsatisfiable as written: GitHub issue bodies and comments live in no file in the tree. (The repository
+is public, so `WebFetch` of `github.com/drmoisan/TaskMaster/issues/<N>` does work — see
+`task-researcher/reference_github_issue_search_without_gh.md` — but only if the prompt says so; a
+prompt that names `gh` gives the agent no route.) On that run the agent produced no artifact at all for a very
+long period after being handed four `gh issue view` instructions.
+
+The consequence is structural, not stylistic. Issue text is very often the ONLY place a root-cause
+history, a premise correction, or a consolidated acceptance-criteria set exists — on #743 the whole
+acceptance-criteria set lived in a consolidation comment, and the fact that the maintainer's first lead
+had been refactored away was only discoverable by reading #592 against the current tree.
+
+**How to apply.** Before delegating research that depends on issue history, run the `gh issue view`
+calls yourself and write the results into a canonical evidence artifact under
+`<FEATURE>/evidence/other/`, then point the research prompt at that FILE rather than at `gh`. There is
+no SendMessage tool, so a prompt that asks for an impossible tool cannot be corrected once launched —
+see [[no-sendmessage-relaunch-with-resume-brief]] and
+[[agent-tool-cannot-course-correct-running-subagent]].
+
 Do not skip these because the delegate sounded careful. Related:
 [[subagent-self-reported-correction-can-be-false]] and
 [[feedback_verify_subagent_capability_claims]]. Note the converse also holds and is worth the same

.claude/agent-memory/prd-feature/MEMORY.md

@@ -32,3 +32,4 @@
 - [500-line ceiling counts TOTAL lines](feature_line_ceiling_counts_total_lines.md)
 - [Negative control must isolate the code fix](feedback_negative_control_must_isolate_the_code_fix.md)
 - [Outcome AC when the mechanism is unverified](feedback_outcome_ac_when_mechanism_unverified.md)
+- [Issue 671: projections-only evidence](project_671_projections_only_evidence.md) — no new raw TRX/coverage XML in the repo (effective 2026-09-12); name fixed-filename Markdown projections in full feature-relative backticked paths, never bare `evidence/...`

.claude/agent-memory/prd-feature/feedback_backticked_paths_are_the_change_footprint.md

@@ -53,6 +53,14 @@ sibling file under a space-free path, that is the reason; record it in the spec'
 do not "restore" the research artifact's original placement. Same mechanism makes the mandated
 CLAUDE.md msbuild command strings safe to quote.
 
+**Evidence paths are the second permitted backtick site (seen on #839, 2026-09-12).** When the caller
+mandates that evidence artifact paths be backticked in full, they are allowed outside the Write Set
+because they fall under the feature-folder `**` glob that is itself a Write Set line; say so in the
+top blockquote. Never backtick a bare repo-root `evidence/...` path (a sibling produced 44 bogus
+blast-radius entries that way). Also unbacktick msbuild switch tokens such as the
+TreatWarningsAsErrors property: a slash-bearing whitespace-free token is harvestable. See
+[[671-projections-only-evidence]].
+
 **The seeded spec template is itself a source of false write claims (seen on #798, 2026-09-07).**
 The promotion scaffold copies `issue.md` prose into Context / Repro & Evidence, and that prose
 arrives with backticks already around paths that are *not* write claims: the debug-log path under

.claude/agent-memory/prd-feature/project_671_projections_only_evidence.md

@@ -0,0 +1,12 @@
+---
+name: 671-projections-only-evidence
+description: Maintainer decision on issue 671 (effective 2026-09-12) - commit Markdown evidence projections only; no new raw test-result XML/TRX or coverage XML in the repo; shapes every spec's Test Strategy and evidence ACs
+metadata:
+  type: project
+---
+
+By maintainer decision on issue 671, effective as of 2026-09-12, feature work commits **projections only**: numeric coverage figures and test-run summaries go into Markdown evidence artifacts under the feature folder's `evidence/<kind>/` tree, and the raw tool output (TRX, Cobertura XML, .coverage) is discarded, never added to the repository.
+
+**Why:** Raw XML evidence bloated the repo and produced meaningless blast-radius entries; the maintainer ruled on #671 that the Markdown projection is the artifact of record.
+
+**How to apply:** In every spec's Test Strategy and Acceptance Criteria, name Markdown projections (Timestamp / Command / EXIT_CODE / output summary) as the evidence, and add an AC that no `.xml`, `.trx` or `.coverage` file is added by the diff. Use fixed, non-timestamped filenames for projections that ACs must name, so no `<timestamp>` placeholder is needed inside a backticked path; the run timestamp goes in the artifact's `Timestamp:` field. Evidence paths must be written feature-relative and in full, e.g. `docs/features/active/<slug>/evidence/regression-testing/<name>.md`, never a bare repo-root `evidence/...` (a sibling item did that on 2026-09-12 and produced 44 meaningless blast-radius entries). Related: [[backticked-paths-are-the-change-footprint]], [[full-bug-spec-only]].