[HumanBean17]

Summary

Batch of five small, verified-still-valid issues from the triage of the open issue list: one real bug, one small CLI contract tightening, one cosmetic refactor, and two docs gaps. Plus housekeeping on two stale issues.

Fixes

• _graph_pass_progress emits status=done even when the pass body raised (done→failed flicker) #459 (bug) — _graph_pass_progress emitted status=done unconditionally in its finally, so a pass whose body raised still reported pass=N/6 status=done → the renderer marked the bar graph ✓ underneath the parent-side failed (done→failed flicker; two terminal events for one failed build). Now tracks whether the body raised and emits status=failed when it did. The renderer already treats failed as terminal for kind=graph, so no renderer change was needed.
• jrag: reject --count + --exists together (argparse mutex) #429 — --count and --exists are now mutually exclusive at parse time (add_mutually_exclusive_group on _common_parser; inherited by every query/listing/traversal subcommand). Both together previously fell through to the silent exists-wins precedence. The usage error flows through the existing clean usage_error envelope (honors --format json, rc=2). --fields stays independent (different axis — documented no-op with the other two).
• refactor(jrag): rename local _emit in _require_kind to stop shadowing the module-level funnel #428 — renamed _require_kind's local _emit closure to _emit_kind_error so it stops shadowing the module-level exit-code/output funnel added in feat(jrag): implement --count/--exists/--fields output flags (#376) #422. Pure rename, no behavior change.
• docs: make operator-doc example paths cross-platform (Windows) #374 — operator docs no longer assume /tmp: docs/JRAG-CLI.md (analyze-pr snippets) and docs/MANUAL-VERIFICATION-CHECKLIST.md now use the relative gitignored tmp/ scratch dir, following the README precedent from feat(platform): declare Windows + macOS support (classifiers, CI matrix, README) #371, with a Git Bash/WSL note for Windows operators.
• docs: document find --fuzzy (identifier loosening) in operator docs #419 — documented find --fuzzy (identifier loosening: exact → prefix → substring on name/FQN) contrasting it with semantic search, in the README flag table (which also gained the --count/--exists/--fields rows it was missing), docs/JRAG-CLI.md § Locate, and the CLI skill/agent artifacts (rules-of-thumb + not_found recovery row). search --fuzzy (a documented no-op) is noted as such.

Housekeeping (done directly, not via this PR)

• Closed jrag find --fuzzy: name prefix/contains fallback (deferred from PR-JRAG-1b) #375 as completed — find --fuzzy shipped in feat(find): implement --fuzzy exact→prefix→contains fallback (#375) #413 (code + tests 15–20 in test_jrag_locate.py verified).
• Closed feat: watch mode — keep the index live as files change #336 as completed — watch shipped as the jrag watch daemon in feat: jrag watch — warm, self-refreshing index daemon #424; residuals tracked in ci: run the heavy jrag-watch tests (golden/COW/warm) in CI or nightly #425/jrag watch: skip client-side graph load when the daemon is alive (transient rc=2 during reindex) #426.

Tests

• New: test_graph_pass_progress_reports_failed_when_pass_body_raises / ..._reports_done_on_clean_exit (direct unit tests of the context manager, incl. exception propagation), test_count_and_exists_together_rejected_at_parse_time (envelope shape + rc + no index access needed).
• render()-level exists-wins precedence test (test_exists_takes_precedence_over_count) intentionally kept: the render precedence chain itself is unchanged defense-in-depth; only the parse-time path is new.
• scripts/sync_agent_artifacts.py run — skill/agent edits are mirrored into install_data (test_install_data_sync green).

Test run

Full suite: 2224 passed, 45 skipped, 2 failed — both failures verified pre-existing on clean master (stash-checked) and unrelated to this diff:

• tests/bench/test_grade.py::test_judge_answer_returns_grade — local claude CLI doesn't know the glm-5.2 judge model (environmental).
• tests/usage/test_cmd_usage.py::test_populated_rollup — IndexError, reproduces on clean master (same family as the test: pre-existing failures in the full heavy suite (lancedb ignore assert 2>2 + 3 test-pollution) #427 pollution list).

(Note for reviewers running locally: 21 tests in the first full run here failed only because this machine's venv shims still pointed at the pre-move ~/Desktop/... path — cocoindex/pip shims repaired locally; nothing in this PR touches that.)

Fixes #459
Fixes #429
Fixes #428
Fixes #374
Fixes #419

[HumanBean17]

Both CI failures were latent test defects this PR's run tripped — master CI last ran 2026-09-13, the day both fixtures were written; not caused by the diff under test. Fixed in f531454:

1. test_populated_rollup (all 3 OSes) — time bomb: the synthetic watch event carried a hardcoded ts=2026-09-13T09:00:00Z, but load_events drops events older than the rollup's 7-day window → the fixture expired 2026-09-20. Now uses rfc3339_now().

2. test_bank_chat_bulk_build_matches_committed_baseline (ubuntu only) — the sampled-edges check ran LIMIT 3 without ORDER BY, riding LadybugDB's storage scan order, which is not stable across platforms — ubuntu sampled a different first-3 (callee_declaring_role SERVICE vs OTHER; the baseline's OTHER,OTHER,OTHER was itself just the generating machine's scan order). Now pins the sample with ORDER BY a.id, b.id (content-keyed ids), compares CALLS roles order-tolerantly, and the fixture's sampled_edges is regenerated accordingly (also drops storage-internal _SRC/_DST/_ID offsets from the anchor — nothing asserted them).

Full local suite after the fix: 2223 passed / 45 skipped; the only remaining local failure is bench/test_grade.py::test_judge_answer_returns_grade, which needs the glm-5.2 judge model registered in the local claude CLI (environmental; CI passed it).