MIN_SAMPLE (added in febf9f6) makes a one-vector space non-anisotropic: one
sample is not a direction. That verdict then fell straight into the `isotropic`
branch of `centroidStatus`, so a space holding a single block reported:
reason = isotropic
Measured before this commit:
1 block reason=isotropic available=false count=1 isAnomalous=false meanLength=1.0000
2 blocks, orthogonal reason=ok available=true count=2 isAnomalous=true meanLength=0.7071
## Why the distinction is the whole point of C5
isotropic this space will never have a common direction — stop expecting centring
too_few_samples write a few more blocks and it starts working
Two diagnoses with opposite remedies. C5 exists so that an operator asking "why is
centring off?" gets an answer they can act on; reporting the first when the
second is true sends them to give up instead of to keep using the thing.
This is my own hole: I added MIN_SAMPLE, and the misclassification exists only
because of it. `zero_sample` (a Go branch) already covered count == 0; nothing
covered 0 < count < MIN_SAMPLE.
## Fix
A `too_few_samples` branch placed BEFORE the isotropic verdict — order is
load-bearing, because the isotropic check is what used to swallow this case.
no_centroid → dim_mismatch → zero_sample → stale → too_few_samples → isotropic → ok
Stale still comes first: a centroid built from one block that has since seen
twelve is stale, and rebuild is the action, not "write more".
## Criterion
`test/phase-b-c5-reasons.test.js` gains case 11, asserting both sides of the
boundary: one vector reports `too_few_samples`, two vectors do not (MIN_SAMPLE is
2, so two samples must clear the branch).
Mutation — deleting the branch:
not ok 11 - C5: a one-vector space reports too_few_samples, not isotropic
error: C5: "not enough blocks yet" and "this space has no common direction" lead
to opposite actions. ... (actual: 'isotropic')
## Still HA-divergent, recorded not fixed
Two orthogonal blocks report `ok` / available with meanLength 0.7071 — i.e. the
0.6 threshold calls a two-sample space anisotropic and centring engages. That is
HA-consistent behaviour (HA has no minimum sample count), and it is now visible
rather than silent because the count and meanLength are in the reason's siblings.
Not changed here: raising the threshold would depart from the port.
Gates: npm test 302/302 (0 skipped), c5 11/11, npm run check 12/12.
jianf
7ad76167c5