[junixlabs]

Everything between v0.2 and a release I would defend.

1. The first outside run produced 22 errors from one misunderstanding (#41)

ids use hyphens, keys use underscores, and the validator printed the raw regex. Fixed with
a sentence, not a rule change:

id: 'flow.my_flow' is not a valid id. ids use hyphens and keys use underscores:
`flow.my-flow`, not `flow.my_flow`, beside `coverage_declared` and `transitions_to`.

2. Five assertions on one subject produced ten errors and an INVALID model (#40)

The check flagged every pair whose claim strings differed. Nothing reads a claim, so
"different string" was standing in for "incompatible proposition". Two findings now,
because two things were conflated:

Finding	Level	Condition
CONTRADICTION	error	same subject, overlapping selectors, and the latest observations disagree — one supported, one refuted
SUBJECT_NOT_DISCRIMINATING	warn	same subject, overlapping selectors, different wording — once per subject

Measured on the reported case: 10 errors and INVALID → 1 warning and VALID. A test
proves the genuine contradiction stays an error, and its positive control fails if the
error is disabled.

3. #5 measured, and the answer surprised me

graph	entities	median reached by one change	worst
synthetic, 10 flows	61	15%	20%
synthetic, 50 flows, fanout 4	301	9%	15%
synthetic, 200 flows, fanout 6	1,201	5%	7%
examples/ecommerce	32	61%	66%
examples/traceos-itself	45	56%	89%

Tiers do not collapse as the graph grows — the hypothesis is refuted at scale. They
collapse on small dense models, which is what a model looks like in its first week. So
impact() reports share_reached and prints DID NOT DISCRIMINATE above 50% — INV-026's
shape one level down, and a fact about the model rather than a finding about the change.

Two corrections to my own measurement before those numbers were trustworthy: a denominator
that counted only flows and nodes while the tiers also return states and externals
(producing an impossible 103%), and a summary that added medians across tiers.

4. #10 measured — the number DIRECTION.md said governs everything

Unit: a re-observation, a reference a change invalidated that someone must open and read.

	n	median	mean	max
before symbol-granular decay	20 changes	11.5	15.0	80
after symbol-granular decay	7 changes	3.0	4.6	14

A change used to force re-reading ~88% of the model's references; it now forces ~19%.
Caveat stated in the issue and the doc: n=7, one repository, one author who wrote both the
model and the code.

5. Two false-decay bugs in the extractor I shipped yesterday

• git log -L reports the range at the last commit that touched the symbol, not at
  HEAD. A 111-line insertion moved validate_intents from 717 to 747 and the hash was
  taken over the wrong 68 lines. Now git blame -L :<symbol> <file>, which answers at HEAD.
• git's funcname block includes the blank lines after a symbol, so adding a function
  after one read as an edit to it. Trailing blanks are dropped before hashing, the same
  rule observed_norm already applies.

Both were false positives in the safe direction, and both were still wrong — a checker that
cries wolf gets discharged by rubber stamp. Sixteen references were re-observed because the
corrected extractor invalidated hashes computed by the buggy one; that is a real re-read,
not a backfill.

Checks

41/41 groups, 196/196 assertions. Every new behaviour falsified: the contradiction
error disabled, the per-pair finding restored, the discrimination flag pinned to a
constant, trailing blanks kept, and git log -L put back — each fails the suite.
check_docs 28/28, check_locators 47/47, ratchet exit 0.