Hunt R-DC6E6F: results

Question. Does scripts/make_context.py --check detect public functions, classes, constants, or module additions under compiler/, and what is the exact sensitivity boundary?

Answer. It does not detect them. Across all 3 Python modules (__init__.py, catalog.py, semantics.py), 10 LLVM IR fixtures, and 1 documentation file in compiler/, every modification, addition, rename, and line-count alteration passes scripts/make_context.py --check completely undetected (0 of 17 curated compiler mutants detected, 0.0% detection rate; 0 of 37 public symbols detected across the exhaustive census, 0.0% detection rate). The generated CONTEXT.md is byte-identical (0 diff lines) because compiler/ is omitted from the scanned directory targets in scripts/make_context.py (PKG, DISCOVERY, HARNESS, DOSSIER, DOCS, SCRIPTS, TESTS) and from build_flat().

Conversely, the guard catches 100% of positive controls in scanned packages (5 of 5 detected, 100.0%), including public symbol additions to zeta/core.py, additions to harness/departments/compiler_department.py, new documents in docs/, new scripts in scripts/, and new test files in tests/.

Reproduce: python3 hunts/r_dc6e6f/probe.py (~15 s, standard library only). Raw data: hunts/r_dc6e6f/results.json.

Controls

All baseline and undo controls ran and held cleanly across the entire run:

control	what it rules out	outcome
Control 1: unmutated sandbox run through guard	catches caused by copy artifacts or tree dirty state	exit 0, `CONTEXT.md is up to date.`
Control 2: sandbox state restored and guard re-run after every mutant	cross-mutant contamination or residue	quiet (exit 0) after all 22 curated mutants and 37 census mutants
Control 3: positive controls in scanned packages	broken or unresponsive guard invocation	5 / 5 fired (100.0% caught with diff > 0)

The curated battery

diff is the number of unified diff lines in regenerated CONTEXT.md. token is whether the mutated symbol's name appears in that diff.

Category 1: Public functions added under compiler/

id	mutant	fired	token in diff
M01	`compiler/catalog.py`: append `def new_catalog_probe(): pass` (`__all__` untouched)	no	no
M02	`compiler/catalog.py`: append `def new_catalog_probe(): pass` (added to `__all__`)	no	no
M03	`compiler/semantics.py`: append `def new_semantics_probe(): pass` (`__all__` untouched)	no	no
M04	`compiler/semantics.py`: append `def new_semantics_probe(): pass` (added to `__all__`)	no	no
M05	`compiler/__init__.py`: append `def new_init_probe(): pass`	no	no

0 / 5 detected (100% blind).

Category 2: Public classes and uppercase constants in compiler/

id	mutant	fired	token in diff
M06	`compiler/catalog.py`: add `class ProbeCatalogClass` (added to `__all__`)	no	no
M07	`compiler/semantics.py`: add `class ProbeSemanticsClass` (added to `__all__`)	no	no
M08	`compiler/semantics.py`: add constant `PROBE_COMPILER_CONSTANT = 100`	no	no

0 / 3 detected (100% blind).

Category 3: In-place modifications and renames in compiler/

id	mutant	fired	token in diff
M09	`compiler/catalog.py`: rename `instruction_count` to `probe_inst_count` in place and in `__all__`	no	no
M10	`compiler/semantics.py`: rename `backend_status` to `probe_backend_status` in place and in `__all__`	no	no
M11	`compiler/catalog.py`: modify docstring first line of `load_ir`	no	no
M12	`compiler/semantics.py`: modify signature of `refinement()` in place	no	no

0 / 4 detected (100% blind).

Category 4: Module, fixture, and documentation additions under compiler/

id	mutant	fired	token in diff
M13	`compiler/optimizer.py`: create new module with public functions and docstring	no	no
M14	`compiler/fixtures/probe_fixture.ll`: create new LLVM IR fixture	no	no
M15	`compiler/FINDINGS.md`: edit title heading and body content	no	no

0 / 3 detected (100% blind).

Category 5: Line-count variations in compiler/

id	mutant	fired	diff	token in diff
M16	`compiler/catalog.py`: append 20 blank lines (pure line count tell)	no	0	no
M17	`compiler/semantics.py`: delete 30 lines (pure line count tell)	no	0	no

0 / 2 detected (100% blind).

Category 6: Positive controls in scanned packages

id	mutant	fired	diff	token in diff
P01	`zeta/core.py`: append public function `probe_zeta_fn` (added to `__all__`)	yes	3	yes
P02	`harness/departments/compiler_department.py`: append public function `probe_dept_fn` (added to `__all__`)	yes	3	yes
P03	`docs/probe_test_doc.md`: create new markdown document	yes	1	yes
P04	`scripts/99_probe_script.py`: create new runnable script	yes	1	yes
P05	`tests/test_probe_case.py`: create new test file	yes	4	yes

5 / 5 detected (100% caught).

Repository census and exhaustive symbol mutation analysis

A full static AST analysis across compiler/ reveals:

Modules under compiler/: 3 Python modules (__init__.py, catalog.py, semantics.py).
compiler/__init__.py: 6 lines, 0 functions, 0 classes, 0 constants (module docstring only).
compiler/catalog.py: 341 lines, 7 public functions, 2 public classes, 5 uppercase constants, declares __all__ with 14 symbols.
compiler/semantics.py: 809 lines, 10 public functions, 5 public classes, 8 uppercase constants, declares __all__ with 21 symbols.
Total public symbols in compiler/: 37 public symbols (17 functions, 7 classes, 13 constants).
Fixtures and docs: 10 LLVM IR fixtures in compiler/fixtures/ (.ll files), 1 document compiler/FINDINGS.md (369 lines).
Exhaustive symbol mutation results:
17 of 17 public functions: 0 / 17 fired (0.0% detection rate, 100% blind).
7 of 7 public classes: 0 / 7 fired (0.0% detection rate, 100% blind).
13 of 13 constants: 0 / 13 fired (0.0% detection rate, 100% blind).
Total symbol census: 0 / 37 fired (0.0% detection rate, 100% blind).

What the numbers say, in order of how much they matter

The blindness is total across the package. scripts/make_context.py defines explicit scan roots: PKG (zeta/), DISCOVERY (ontology/), HARNESS (harness/), DOSSIER (dossier/), DOCS (docs/), SCRIPTS (scripts/), and TESTS (tests/). Because compiler/ is omitted from these roots, every file under compiler/ is entirely invisible to scripts/make_context.py --check.
Line-count tells do not operate on unscanned packages. In scanned packages, any line count alteration trips the per-module line count in CONTEXT.md (*NNN lines*). In compiler/, neither line counts nor symbol signatures are extracted, so adding, deleting, or reflowing lines in compiler/ produces zero diff in CONTEXT.md.
The wrapper department is guarded, but the subject implementation is not. While harness/departments/compiler_department.py (21 exports) and tests/test_compiler_candidate.py (32 tests) are fully indexed and guarded in CONTEXT.md, the actual execution engine in compiler/semantics.py (809 lines, 21 exports) and catalog in compiler/catalog.py (341 lines, 14 exports) have zero presence in the index.
Exclusion from build_flat(). scripts/make_context.py --flat concatenates repository source into CONTEXT_FLAT.md. The group definitions in build_flat() include Package, Discovery layer, Harness, Dossiers, Scripts, Docs, and Tests, but omit compiler/. Thus, prompt concatenation also completely omits compiler/.

What I chose, and why

Isolated sandbox execution: Mutating a throwaway copy under /tmp prevented any dirty working tree state or residual modifications.
Combined curated battery and exhaustive symbol census: The 22 curated mutants test additions, renames, docstring modifications, signature changes, new modules, new fixtures, and line count variations, while the 37-symbol census verifies every individual public symbol in compiler/.
Diff verification with token matching: Checking diff line counts and token presence in CONTEXT.md confirms byte-for-byte identity of the regenerated artifact across all compiler mutations.

What I could not settle

Whether compiler/ should be indexed in CONTEXT.md: compiler/ is a candidate department (department #3) with provisional findings. Whether candidate departments belong in the primary CONTEXT.md index alongside zeta/ and harness/ is an architectural design decision for repository maintainers.

Loose threads

Other top-level packages unscanned by make_context.py: meta/ (B01) is similarly unscanned. In addition, interactive_lab/, publication/, and mathlib-wishlist/ are top-level directories not indexed in CONTEXT.md.
Why it might matter: New contributors or agents reading CONTEXT.md have no index visibility into tools or metadata in these directories.
First step: Review the top-level directory census and decide whether a "Tooling and metadata" section should be added to scripts/make_context.py.
Exclusion of compiler/ from CONTEXT_FLAT.md: When --flat is used to dump the repository for single-prompt contexts, compiler/ is omitted entirely.
Why it might matter: An agent given only CONTEXT_FLAT.md will be unaware of compiler semantics and catalog definitions.
First step: Add ("Compiler", sorted((ROOT / "compiler").glob("*.py"))) to build_flat() in scripts/make_context.py if candidate departments are intended to be included in full dumps.