WHY: test_acquire_blocks_and_returns took t0 after the drain, so the cell timed
only the part of the 0.05 s refill that happened after the scheduler returned to
it. Observed red in a loaded full-suite run at 7c36185:
'assert 0.03909331373870373 >= 0.04' -- 10.9 ms of untimed prefix. Deterministic
repro: pre-fix cell + 11 ms injected gap -> 0.03897/0.03912, rc 1.
FIX: t0 now precedes try_acquire(1.0), which is where the bucket sets _last, so
dt is >= 0.05 by construction (acquire cannot return before the refill origin is
1/rate old) and the 0.04 floor has ~10 ms of margin instead of 1 ms.
MEASURED: repo cell 9 passed; pre control rc 1 dt 0.03897-0.03912; post control
rc 0; same-gap in-process dt pre 0.039119 / post 0.050253; non-blocking-acquire
mutant rc 1 (2 failed: blocks + timeout cells); controls/post 15/15 green under
4-CPU saturation; bash wizard-work/beam/CHECK rc 0 twice, 423 passed / 2007
subtests.
CHECK: python3 /root/.hermes/wizard-outputs/volt/2026-10-28-q122/verify.py
wizard-work/beam/test_jsonl_validate.py (new, 25 tests) pins the module's own
docstring contract as process exit codes: 0 valid, 1 bad rows, 2 usage/IO/schema/
decode. wizard-work/beam/jsonl_validate.py repaired in three places:
- undecodable input (and schema) now returns 2 with an `input error:` line and
no summary line, instead of a UnicodeDecodeError traceback and rc 1;
- utf-8-sig on both files, so a leading BOM is stripped not graded as bad data;
- a required field present as null yields one error, not "missing required
field" plus "null not allowed".
Measured: the new test file against the shipped module (md5 f53b6288...) fails
6 of 25 (3 decode, 2 BOM, 1 null); against the repair, 25/25 pass. bash CHECK:
rc 0 with Ran 243 -> Ran 268; with the shipped module swapped back in, rc 1
FAILED (failures=6) -- the gate was rc 0 and blind before. No edit to CHECK:
the file is a unittest.TestCase, so the existing -p 'test_*.py' leg collects it.
The 242-checks line is toposort's import-time counter and does not move (Q105).
verify.py --repo . reads 8 legs, 6 pass, 2 fault (checkin, path-claims) here and
identically at the parent cd5f5c9. path_claim_check's tolerated NOT_ADDED row
for the new file reads CLEARED on this tree. Q105's 27 uncollected files / 131
tests remain open and untouched.
The eight test files Q98 leaves open (plus test_bfs.py, added at 7439cf0 with the
same lock) now resolve their module from pathlib.Path(__file__).resolve().parent
instead of /root/.hermes/wizard-outputs/<lane>.
tools/import_resolve_check.py --repo . at 7439cf0: 95 files, 9 faults, rc 1
after: 95 files, 0 faults, rc 0
Behaviour-preserving: every host copy the tests were reading is byte-identical
(md5sum) to the repo copy. Three absolute literals in spark/test_validator.py
that the static leg does not grade (load_json(...) argument strings at pre-fix
lines 12/16/25) are re-pointed too.
ledger/test_scratch_manifest.py was red before any edit (str ROOT vs
gen_manifest ROOT.iterdir), would have written the tracked
ledger/wizardwork-file-manifest.tsv once that was fixed (MANIFEST derived from
OUT at import time), and asserted nothing. It now re-points ROOT/OUT/MANIFEST,
guards MANIFEST inside a tempfile.mkdtemp(), and asserts the exact manifest row.
Non-vacuity: scratch/q98b/poison_probe.py copies each lane, runs the test from a
foreign cwd (rc 0), poisons the module inside the copy, and re-runs -- 8/8 go
rc 1, so none of them still reads the host.
Lane gates unchanged: beam CHECK rc 0, spark CHECK rc 0 (VERIFIED), ledger CHECK
rc 1 on both sides (commit-census drift, red at the parent commit too).
Q99's subject was a single tracked name holding seven newline bytes,
`wizard-work/beam/CHECK\ncontent">\n#!/bin/bash\n...`, added by accident at
`a46547c` (a shell-quoting accident, printf/heredoc output used as a pathspec),
218 bytes, blob `763936956721f3ba047d25b884f17862469482c3`.
Before this commit at `47532eb`: `git ls-files -z` yields 3,816 names and seven
of those bytes are `\n`, so a newline-splitting consumer reads 3,823 rows for
3,816 entries. The one path does not merely inflate the total, it impersonates
files: `git ls-tree -r -z --name-only HEAD | tr '\0' '\n'` prints it as eight
consecutive rows (lines 670-677 at `47532eb`), the first four of which are
`wizard-work/beam/CHECK`, `content">`, `#!/bin/bash` and
`# CHECK: verify all beam reliability primitives` -- four names a per-file census
or a per-file gate would try to open.
Content is not lost and nothing is being decided about it: the blob is
byte-for-byte `wizard-work/beam/CHECK` minus its final newline, verified against
`a46547c:wizard-work/beam/CHECK` and against HEAD's copy (`bad == good[:-1]`,
185 B vs 186 B, no differing offset), and the real CHECK is untouched. The blob
remains reachable from `a46547c`, so `git cat-file blob 7639369` still resolves
after this commit.
Guards run before the unlink: the worktree bytes were asserted equal to the blob
the index pinned, and that blob was asserted reachable from a ref.
After this commit: tracked 3,816 -> 3,815, names containing `\n` 1 -> 0,
newline-split rows 3,823 -> 3,815 (== entry count). An index-derived census keyed
on backslash-escaped names (`q99_census.py`, scratch) reads before=3816
after=3815, removed=1, added=0, changed=0 -- the removal is the whole delta, and
Q99's done-means ("a census run before and after differs only by that entry") is
met under a key that can express the entry at all.
Gate A/B on a scratch clone, committed there so HEAD moved: `tools/verify.py`
reads 5 pass / 3 fault on both sides with the same three legs (path-claims,
selftest-sweep, write-axis) and the same counts, so the removal regresses no leg.
The two path-claims faults are the `inbox/` rows and are untouched by this
commit.
Probes (read-only, outside the repo, under the volt profile cache):
/root/.hermes/profiles/volt/cache/scratch/q99/
newpath_probe.py, pin_content.py, q99_census.py, q99_worktree_state.py,
q99_apply_live.py, census_before.tsv, census_after.tsv, verify_after.txt
Rebuild test_no_path_compression_find_stable on 4 nodes so the node it asks
about (3) sits two levels down under union-by-size; the old cell built
union(0,1);union(1,2) -> parent [0,0,0,0], so find(2) rewrote nothing and a
path-compressing find passed the cell that names the mechanism.
New test_refuses_out_of_range_index and test_refuses_non_int_index grade the
_check_index contract e33aa30 added and no cell graded: every entry point that
takes an index refuses -1/n/10 and bool/float/None, and a refused call leaves
parent, sz, history and components unchanged.
Docstring API block now states the raise contract instead of the pre-e33aa30
signatures.
Measured (scratch copies, mutants built from this head):
M-GUARD (6 call sites deleted) HEAD cells rc 0 -> these cells rc 1
M-SIZE still rc 1 on test_shape_union_by_size
M-COMPRESS (plain) HEAD only test_rollback_restores, now the
compression cell too
journalled compression HEAD 12 cells all pass; only the rebuilt
compression cell fails
Lane: bash wizard-work/beam/CHECK rc 0, Ran 232 (230 + 2).
- the docstring said "RFC 8785-ish"; it now says NOT RFC 8785 (JCS) in those
words, because the hedge only holds if the three readings are written down.
- measured at this file (scratch/q82/floats.py, output in the closure record):
-0.0 -> "-0.0" where RFC 8785 Table 1 gives "0"; 1e-06 -> "1e-06" where the
table gives "0.000001"; 9007199254740993 -> "9007199254740993" where JCS's
double-based maximum positive integer is 9007199254740992.
- the docstring also states that sort_keys sorts by code point, not by UTF-16
code unit, so key order diverges from JCS above U+FFFF.
- no behaviour change: the seven cells still pass and `bash wizard-work/beam/CHECK`
is rc 0, Ran 198, the same number as the previous commit.
CHECK: bash wizard-work/beam/CHECK | grep -q '^Ran 198 tests' && python3 -m unittest discover -s wizard-work/beam -p test_json_canonical.py
- the file was seven `test_*` functions that nothing called, plus a module-level
`print("ALL PASS")`: running it printed a verdict with zero assertions, and
`python3 -m unittest discover -s wizard-work/beam -p test_json_canonical.py`
collected 0 (rc 5, NO TESTS RAN) while the import still printed ALL PASS.
- the seven cells are now methods of one `unittest.TestCase`; discover collects
7, `python3 wizard-work/beam/test_json_canonical.py` runs 7 OK, and
`bash wizard-work/beam/CHECK` moves Ran 191 -> Ran 198, rc 0, OK.
- the ALL PASS print is deleted; there is no module-level verdict on import.
- cells bite, measured against a mutant battery (harness
/root/.hermes/profiles/volt/cache/scratch/q82/battery.py, rows in
mutant_results.json, one tree per mutant under .../q82/mut/). Control 0/7 dead;
`sort_keys=True` -> False: 2 dead (test_basic, test_nested); `allow_nan` allowed
with `_normalize`'s non-finite guard removed: 1 dead (test_reject_nan);
`ensure_ascii=True`: 1 dead (test_unicode); non-string-key guard removed: 1 dead
(test_reject_nonstr_key); `canonical_bytes` returning str: 3 dead
(test_bytes, test_roundtrip, test_unicode); bool serialised as "true"/"false":
2 dead (test_basic, test_roundtrip). All 7 cells have a named killer.
- one equivalence measured, not counted as a kill: `allow_nan=True` alone is 0/7
dead, because `_normalize` raises CanonicalError on a non-finite float before
`json.dumps` is reached. The `allow_nan=False` argument is defence in depth on
this API, not a reachable behaviour; the mutant that removes the guard is what
a cell can see.
CHECK: bash wizard-work/beam/CHECK
Seventeen unittest.TestCase cells, each driving the committed module through
the CLI in a subprocess so the exit code is what is graded. MOD resolves from
this file's own directory, not a host path.
Negative cells Q79 asked for, all present: a directory symlink added after
make (rc 1, EXTRA), one removed (rc 1, MISSING), a manifest stored inside the
root (clean, and make twice is byte-identical), a missing root (rc 2, no
artifact), a root that is a file (rc 2), six corrupt-manifest shapes and an
absent manifest (rc 2), plus graded faults confined to rc 1.
Bite measured, not asserted: the same suite against edac5cf's manifest.py is
11 of 17 failing; against a copy whose check() body is `return []`, 9 of 17
fail, including tampered/extra/missing -- the case Q79 records as passing the
lane gate. `bash wizard-work/beam/CHECK` -> Ran 191 tests, OK, rc 0.
CHECK: python3 wizard-work/beam/test_manifest.py
Q79's done-means allows either retiring the overlapping surface or naming
which file is canonical, not both. Retiring sha256_manifest.py would break
test_sha256_manifest.py and the reply-01..08 history that grades it, so this
names the split instead: the JSON directory manifest (sizes, symlink targets,
directory symlinks, 0/1/2 contract) is manifest.py; the plain
`digest path` text shape stays in sha256_manifest.py. One docstring, no
behaviour change.
CHECK: python3 -m unittest discover -s wizard-work/beam -p 'test_sha256*.py' -q
- rc 2 for input that cannot be graded: a root that is missing or is not a
directory, an unreadable or corrupt manifest (bad JSON, no entries list, a
malformed entry, an unknown type), and a manifest that cannot be written.
rc 1 is reserved for a graded fault, rc 0 for a clean check.
- make on a missing root now writes nothing and exits 2; before, it wrote a
0-entry manifest and read green.
- _walk records a symlinked directory as a `symlink` entry and does not
descend it, so adding or removing one after make is a graded difference
(EXTRA / MISSING, rc 1) instead of rc 0.
- a manifest stored inside its own root is skipped in make and check by
relpath, so make twice is byte-identical and check reads OK.
- a root it cannot read (OSError from os.walk or open) is Ungradable, rc 2,
instead of an uncaught traceback that exited 1.
CHECK: python3 wizard-work/beam/manifest.py make /root/troupe-standup/wizard-work/beam /tmp/q79-manifest.json && python3 wizard-work/beam/manifest.py check /root/troupe-standup/wizard-work/beam /tmp/q79-manifest.json
test_sliding_window_median.py: the seven cells were module-level functions
behind `if __name__ == "__main__"`, so `unittest discover` collected none of
them -- `python3 -m unittest test_sliding_window_median` printed `Ran 0
tests` (rc 5 on Python 3.14) and the lane gate stayed at `Ran 151` while a
module with a floored even-median passed it at rc 0. They are TestCase
methods now, with the same seven assertions, the module resolved from this
file's directory (Q61 rule), and one added cell, test_cell_census, which
pins the count at 8 so a deleted cell is a red gate.
test_robin_hood_hash.py: `sys.path.insert(0, "/root/.hermes/wizard-outputs
/beam")` -> `str(pathlib.Path(__file__).resolve().parent)`. That host insert
is what made CHECK rc 1 unmasked with `ImportError: 'test_safe_eval' module
incorrectly imported from /root/.hermes/wizard-outputs/beam`.
Measured at this commit: `bash wizard-work/beam/CHECK` rc 0, `Ran 159 tests`,
OK -- unmasked, with /root/.hermes/wizard-outputs masked to an empty tmpfs,
and in a copied lane with no .git. 151 + 8 = 159. Mutants: 8 of 9 killed
(M8 is an equivalent mutant, 0 differences in 20,000 batch and 5,000
incremental cases); with the old test file and a floored even-median the gate
is rc 0, Ran 151, OK. Record: wizard-work/volt/q76-2026-10-26.md.
beam
ledger
volt
Timmy
quill
spark