[Deicyde]

A follow-up to #482: the skeleton, impact and library index probes get their records-file contract from one lean_probe helper instead of filling it in three places.

Change

• One helper. lean_probe.render_probe_template(name, **fields) loads probes/<name>, fills fields, and adds the two contract fields every probe template uses: output_env (the variable naming the records file, which run_probe sets) and output_limit (the byte cap _read_probe_records enforces). It is in lean_probe.__all__, since impact and library.generate import it.
• Read at call time. The helper reads PROBE_OUTPUT_ENV and DEFAULT_PROBE_OUTPUT_LIMIT when it runs, not as default arguments, so tests that patch autoform_cli.lean_probe.DEFAULT_PROBE_OUTPUT_LIMIT still reach the rendered program. A caller that passes either field itself gets a TypeError.
• Three callers. _render_probe_helper, render_impact_probe and render_library_probe call the helper and drop their two contract lines. lean_probe._probe_template, impact._impact_template and generate.py's inline template load are deleted.
• Template headers. The header comments of skeleton_probe.lean, impact_probe.lean and library_probe.lean now say the fields are filled through lean_probe.render_probe_template, which adds output_env and output_limit. The templates are format strings, so the comments spell the field names without braces.
• One test. test_every_probe_template_gets_the_records_file_contract renders all three probes with the limit patched to 1234 and checks that each reads the records-file variable and declares probeOutputLimit as 1234. It passes on main too, since main's three renderers already read the limit at call time; it is there to fail if the helper ever binds the limit as a default argument.

No command, flag, JSON key or message changes, so there is no CHANGELOG entry.

Net: +45/−22 across 7 files.

Verification

• Rendering the three probes on main and on this branch (pure string rendering, no Lean) gives 803/216/181 lines on main and 805/218/183 here. diff -u shows only the edited header comment lines; every other line is identical.
• Replacing the helper with a version that binds the limit as a default argument makes all three cases of the new test fail.
• When called directly on macOS, the new test's three cases pass, as do the existing pins on the rendered probes: the helper's probeOutputLimit line in tests/test_lean_probe.py, the records-file variable in tests/test_impact.py, and no unfilled output_limit field in tests/test_library_generate.py. So does the fake-Lean records-file limit test; the real-Lean one skips there.
• The locked ruff 0.15.17 check passes on the touched Python files. ruff format --check reports the same findings on them as on the base; CI does not run the formatter.
• No file in the tree names _probe_template or _impact_template.
• CI coverage:
  • test (Python 3.10 and 3.13) runs ruff check autoform_cli servers tests and pytest -q, which runs the new test and the render pins; none needs Lean.
  • real-lean runs tests/test_impact.py and tests/test_library_generate.py against real Lean, so the rendered impact and library probes, header comments included, are run by Lean.
  • real-lean-skeleton runs tests/test_lean_probe.py against real Lean. test_probe_stops_before_its_records_file_exceeds_the_limit patches the limit to 1024, builds the helper module from the rendered skeleton template, and checks that the probe stops before its records file passes that size.
  • windows-hardening runs only the node IDs it lists, none of which this branch changes.

Overlap

• Based on main after Move the lean_probe runner tests into tests/test_lean_probe.py #504 (which moved the runner tests into tests/test_lean_probe.py, where the new test goes) and Refuse a lean_lib that is not an array of named tables, and tidy lean_probe #505 merged; it merges cleanly with current main.
• Claim a source file in import-impact only for a module Lake can build #500 changes lean_probe.py (including __all__) and impact.py, Give the managed-output helpers their own module #501 changes lean_probe.py and library/generate.py, Name files by path in input and output errors, and correct the path-free claim #506 changes lean_probe.py's snapshot messages, and Import impact own helpers #492 changes impact.py. At their current heads, each merges with this branch without conflict, and ruff check passes on the touched files of each merged tree.

[Deicyde]

Review at 0e1b812: merge.

Nothing blocks merge, and there are no notes.

• Output is unchanged apart from the comments. I rendered the skeleton helper, an impact probe (with selected modules, module evidence and an artifact directory) and a library probe on main 8155f23 and on this head. The only differences are the three edited header comments, 12 lines in all. No caller passes output_env or output_limit any more, and nothing in the tree still names _probe_template or _impact_template.
• The limit is read at call time. The new test patches DEFAULT_PROBE_OUTPUT_LIMIT and finds the patched value in all three rendered probes. So the existing tests that patch the limit, such as test_probe_stops_before_its_records_file_exceeds_the_limit, still reach the program Lean runs.
• audit_probe.lean is not a format template and has no records file, so leaving it out is right.

Checks

• At this head, tests/test_lean_probe.py, tests/test_impact.py, tests/test_library_generate.py and tests/test_skeleton.py give 478 passed, 74 skipped (real Lean) and 2 failed. One failure is the Python 3.14 PurePath.suffix quirk that also fails on main. The other is test_bounded_command_termination_signal_kills_the_process_group[term], which failed at load average 300 and then passed in 2 of 3 reruns and a fourth. This PR doesn't touch the runner it tests.
• The PR merges cleanly with main and with each of Import impact own helpers #492, Claim a source file in import-impact only for a module Lake can build #500, Give the managed-output helpers their own module #501, Move the LSP token scan into servers/_process.py and drop the underscore from its shared names #502, Name files by path in input and output errors, and correct the path-free claim #506, Index structure and class fields as their projections #509, Stop a Lake or Lean command that fails just as a signal arrives #510 and Remove size caps on published data #512.
• CI passes 9 of 9. The real-Lean jobs run the rendered impact, library and skeleton probes.

Posted by Claude Code (session jackmccarthy-eb).