fix(ci): CodeLens CI/Quality Gate broken on main since #195 consolidation

[Wolfvin]

Mengapa

Ditemukan saat navigasi PR #234 (agent-tooling audit session), bukan scope task itu. Verified CI is red on main itself — checked gh run list --branch main: the merge commit for PR #233 (the last merge before this) already shows CodeLens CI and CodeLens Quality Gate as failure. Not caused by any specific recent PR; pre-existing workflow rot.

Two distinct root causes found (both need separate fixes):

1. quality-gate job calls a dropped top-level command:

Run python3 scripts/codelens.py init .
codelens.py: error: argument {api-map,audit,context,deps,doctor,graph,history,impact,scan,search,security,summary}: invalid choice: 'init'

init was absorbed into the scan umbrella command in the #195 consolidation (scan absorbs: scan/init/rescan — see CONTEXT.md) and no longer exists as a standalone CLI command. The quality-gate workflow YAML (.github/workflows/*.yml, whichever job runs "Initialize CodeLens") still invokes the old standalone form. This fails on every push to main, every PR — the CI has been broken since #195/#196 landed.

2. Test Suite job segfaults in test_large_file_parsing.py:

Fatal Python error: Segmentation fault
  File ".../scripts/parsers/python_parser.py", line 301 in _extract_references_impl
  File ".../scripts/parsers/python_parser.py", line 126 in extract_references
  File ".../tests/test_large_file_parsing.py", line 119 in test_file_above_old_threshold_is_parsed

Crashes the whole pytest process (exit 139), aborting every test after it alphabetically (everything from test_large_file_parsing.py onward never runs in CI, including test_query_graph.py, test_codelens.py variants that sort after it). This matches why the previous session's documented test baseline methodology explicitly used --ignore=tests/test_large_file_parsing.py — it's already known-unstable locally, just never fixed at the source.

Konteks

Both failures reproduce on a clean main checkout, independent of any feature branch. This blocks CI green-checks on literally every future PR until fixed — not something any single feature PR should be expected to fix as a side effect.

Tujuan

CI passes (or at minimum, doesn't segfault/hard-fail) on a clean main checkout.

Constraint

• Per this repo's CLAUDE.md: CI/workflow YAML changes are always done directly by BOS (Claude), not delegated to z.ai workers — historically unreliable for workflow syntax/env-var context.
• Fix feat: add codelens-watch.py — file watcher with auto outline.json generation #1 (quality-gate) is likely a 1-line change (swap init for whatever scan's equivalent init-only invocation is, or drop the step if scan auto-inits).
• Fix refactor: refocus CodeLens — archive frontend, remove garbage, keep core only #2 (segfault) needs actual root-cause investigation in python_parser.py::_extract_references_impl under large-file input — do not just skip/ignore the test without understanding why a C-extension-backed tree-sitter call segfaults (could indicate a real memory-safety bug reachable from real large Python files, not just a test artifact).

Definition of Done

• quality-gate job's "Initialize CodeLens" step succeeds on a clean checkout
• Test Suite job no longer segfaults; either the root cause in _extract_references_impl is fixed, or (if genuinely a test-only artifact unrelated to real usage) documented why with a narrow, justified skip — not a blanket ignore
• A full CI run on main (or a trivial PR) goes green

[Wolfvin]

Reopening: PR #262 fixed the YAML issues (init→scan, check→audit, benchmark) but the segfault fix was MISDIAGNOSED. ulimit -s unlimited does NOT help — the CI log shows the crash is a native use-after-free in tree_sitter_python._binding (not a Python stack overflow), triggered by a shallow 250-line file (test_file_above_old_threshold_is_parsed), NOT the deeply-nested one.

Real root cause: tree-sitter core 0.26.0 (CI, unpinned via tree-sitter>=0.21.0) vs 0.25.2 (local, passes). python_parser.py's keep_alive mitigation is insufficient for 0.26's Node-lifetime change. Fix: pin tree-sitter<0.26. Follow-up issue will be filed for full 0.26 compatibility.