Fix: QuickFiler search box loses keyboard focus when the folder drop-down auto-opens (#680)

.claude/agent-memory/atomic-executor/MEMORY.md

@@ -37,6 +37,7 @@
 - ["Make the citation exist" deltas propagate false facts](project_preflight_citation_match_propagates_false_fact.md) — a fix that copies A's prose into B corroborates A's error in two documents; also, an epic child's issue.md promise to siblings needs a C-constraint AND a verification task
 - [Check-off cites an artifact a LATER task writes](project_preflight_checkoff_cites_later_task_artifact.md) — unsatisfiable in plan order; also covers a phase-5 criterion needing an issue phase-7 opens; fix by body swap, not renumbering
 - ["Skip the pointless drain" note makes a negative test vacuous](project_preflight_drain_scope_optimization_note_makes_test_vacuous.md) — a cosmetic "path X is synchronous, don't drain" note is a behavioral claim; a synchronous callee says nothing about how its caller is reached
+- [Sanitisation task cannot sweep its own record](project_sanitisation_task_cannot_sweep_its_own_record.md) — exactly one residual path always remains; check the record's content constraint, not the scope
 - [Conjunctive criteria break the one-artifact citation rule](project_preflight_conjunctive_criterion_citation_gap.md) — "cites exactly one artifact" makes every AC that conjoins facts from two command steps uncheckable; sweep for the second conjunct ("passes unmodified", "anywhere", "either figure")
 ## Build / toolchain environment
 - [pwsh/git/gh CLI gotchas](project_pwsh_git_gh_cli_gotchas.md) — no jq; pwsh won't concatenate `$(git merge-base ...)..HEAD`
@@ -62,6 +63,7 @@
 - [Evidence <TS> collision clobbers artifacts](project_evidence_timestamp_collision_clobbers_artifacts.md) — same-day remediation overwrites
 - [csharpier pipe-files is a non-enforcing gate](project_csharpier_pipefiles_nonenforcing_gate.md) — use `csharpier check`/`format`
 - [Count-idiom pitfalls: csharpier + Measure-Object](project_count_idiom_pitfalls_csharpier_and_measureobject.md) — "Formatted N" is a PROCESSED count
+- [New .cs files guarantee a format-loop restart](project_new_cs_files_guarantee_a_format_loop_restart.md) — Write emits LF; budget two passes, not one
 - [PowerShell new files need UTF-8 BOM](powershell-bom-required.md) — PSUseBOMForUnicodeEncodedFile
 - [poshqc test MCP carries no verdict/numbers](project_poshqc_pester_mcp_exit_minus1.md) — no exit code, so count gates can't run
 - [poshqc analyze exits 1 on a Warning](project_poshqc_analyze_exit1_on_warning.md) — "EXIT 0 + zero errors" is self-contradictory

.claude/agent-memory/atomic-executor/project_new_cs_files_guarantee_a_format_loop_restart.md

@@ -0,0 +1,32 @@
+---
+name: new-cs-files-guarantee-a-format-loop-restart
+description: Any plan that creates a new .cs file with the Write tool will fail its first csharpier check, because Write emits LF and csharpier demands CRLF — budget one mandatory final-QA loop restart rather than treating it as drift
+metadata:
+  type: project
+---
+
+A C# file created with the Write tool lands with LF line endings. CSharpier reports it as
+`Was not formatted. The file contained different line endings than formatting it would result in.`
+So the FIRST pass of a final-QA loop over any plan that adds a new `.cs` file is always a repairing
+pass, and the loop always restarts at least once. Plan for two passes, not one.
+
+**Why:** on 2026-08-28 (feature 680, `[P6-T1]`) the pass-1 `PRE_FORMAT_CHECK_EXIT` was 1 on three
+files — the two newly Written test files for line endings only, and one Edited file for line endings
+plus a member-chain wrap. Pass 2 was clean. The restart is mechanical and expected; it is not
+pre-existing drift and it does not mean the baseline format check was wrong. Feature 680's P0-T7
+baseline check had exited 0 over 1554 files.
+
+**How to apply:**
+- Treat the first-pass `PRE_FORMAT_CHECK_EXIT: 1` as routine when every reported file is inside the
+  plan's own edited/created set. Record it in the P6-T1 artifact as a non-final pass, restart the
+  loop, and do not investigate it as drift. Only a reported file OUTSIDE that set means pre-existing
+  drift entered the pass.
+- `git status --porcelain` before/after the format command CANNOT observe this. Those files are
+  already `M` or `??` and no commit exists yet, so a content rewrite leaves porcelain byte-identical.
+  `PRE_FORMAT_CHECK_EXIT` is the only discriminator; the porcelain pair only detects a path
+  ENTERING or LEAVING the changed set.
+- The formatter's own `Formatted N files in Xms.` line is a PROCESSED count printed on every run,
+  including a run that rewrote nothing. It never discriminates. See
+  [[count-idiom-pitfalls-csharpier-and-measureobject]].
+- Running `csharpier format` immediately after Writing each new file — before the final QA loop —
+  avoids the restart entirely, but only if the plan does not gate on a pristine first pass.

.claude/agent-memory/atomic-executor/project_sanitisation_task_cannot_sweep_its_own_record.md

@@ -0,0 +1,32 @@
+---
+name: sanitisation-task-cannot-sweep-its-own-record
+description: A "sanitise every committed artifact" task always leaves exactly one path unswept — its own record — so validate the record's content constraint, not the sweep scope
+metadata:
+  type: project
+---
+
+A plan task that sanitises host identifiers out of "every path the close-out commit will
+touch" can never cover its own evidence record, because the record is written after the
+sweep enumerates. Broadening the sweep scope does not close this; the fix is a content
+constraint on the record itself (e.g. "lists only AFTER values — quoting a BEFORE value
+reintroduces the identifier into a committed file").
+
+**Why:** Verified on issue #680 preflight round 5. `P7-T3` swept the union of
+`<FEATURE>/**` and `git status --porcelain`, which provably covers everything `P7-T4`
+commits *except* `<FEATURE>/evidence/other/trx-sanitisation.<ts>.md`. Chasing the scope
+one level deeper is an infinite regress; the plan already carried the right mitigation.
+
+**How to apply:** When validating a sanitisation/redaction gate, enumerate the later
+commit set against the sweep set and expect exactly one residual — the gate's own record.
+Confirm the plan constrains that record's *content* rather than demanding it be swept.
+Two mechanical checks that make the scope claim decidable:
+- `git status --porcelain` collapses untracked directories, so a "no path *name* contains
+  the literal" check over that half is blind to nested names. It is only safe when the
+  collapsed directories also fall under an independently file-enumerated half (here
+  `<FEATURE>/**`). Tracked-but-modified trees like `.claude/agent-memory/**` list per-file
+  and are not affected.
+- Plan check-offs written after the sweep (e.g. the mandatory pre-commit `[x]` flip of the
+  final task) touch an already-swept file but add no identifier, so they are benign.
+
+See [[project_trx_sanitisation_must_be_case_insensitive]] and
+[[project_plan_checkoff_fixpoint_breaks_terminal_clean_tree_gate]].

.claude/agent-memory/atomic-executor/project_trx_sanitisation_must_be_case_insensitive.md

@@ -56,6 +56,25 @@ Verify with a strict parser (`xml.etree.ElementTree.parse`), not with `[xml]` in
 confirm the decoded attribute value: `storage` should read `<repo-root>\...`, and the
 `<UnitTestResult>` count should equal the run's test total.
 
+### Catch the raw `<repo-root>` spelling at PREFLIGHT, not at execution
+
+The escaping fix above is unusable once execution reaches the sanitisation task, because that task is
+normally the second-to-last in the plan and the close-out task immediately after it forbids writing
+anything to disk after the commit. An executor that reads this memory only when it reaches the
+sanitisation task has already lost the chance to apply it without violating the clean-tree gate.
+
+**Why:** on 2026-08-28 (feature 680, `[P7-T3]`) the plan text said "substitute ... the worktree-root
+prefix -> `<repo-root>`" with the placeholder written raw. Following it literally produced five TRX
+files that are no longer well-formed XML wherever a substituted path sat in an attribute value. The
+task's own acceptance is three zero-hit greps plus a zero name-hit count, and all four passed, so
+nothing in the task could detect it. The defect was noticed only after `[P7-T4]` had committed, at
+which point amending would have dirtied the tree the close-out gate requires clean.
+
+**How to apply:** during preflight, treat a sanitisation task that quotes a placeholder containing a
+raw `<` or `>` and targets an XML-shaped artifact (TRX, coverage XML, `.csproj`) as a required plan
+delta: the plan must say `&lt;repo-root&gt;`, or must state that the artifact is not required to
+remain parseable. Raise it as a preflight revision, not as an execution-time judgment call.
+
 ## `/EnableCodeCoverage` drops a host-named binary INTO the evidence tree
 
 A vstest run that combines `/EnableCodeCoverage` with an explicit `/ResultsDirectory:` pointing at