Hesitations

Every moment while using the product where I stopped, doubted, or could not proceed. Each carries a reproduction. Findings here outrank tests: the two worst defects this project has shipped were both found by opening an unfamiliar repository, and neither would have been reached by reasoning.

Truth failures are repaired immediately. Inconveniences are recorded and left.

Repaired

H1 — "refusals: none" on a repository with 134 imports · TRUTH

ky reported nothing refused. Two causes, compounding:

• .js specifiers were not resolved. TypeScript's NodeNext convention writes the emitted extension, so correct TS imports ./x.js where x.ts exists. Zero imports were extracted — no wrong answer, no answer.
• The refusal was conditional on imports.length > 0. A refusal is a claim about what the vocabulary can hold, true whether a run saw five hundred instances or none. Conditional on the count, "we found none" and "we cannot represent these" both rendered as silence.

A refusal that disappears when extraction fails is the failure hiding its own evidence. Repro: acquireRepositoryWorld on any NodeNext TS repo; expect a non-zero import dependency refusal.

H2 — date-fns could not be opened at all · TRUTH

RangeError: Maximum call stack size exceeded, 1631 source files. Math.max(...xs) passes every element as a separate argument and throws past ~10⁴–10⁵. coChange builds up to n²/2 pairs, so the spread blew the stack before a single building existed. Ten sites across src/; symmetrise spread the entire distance matrix.

Repaired with lib/world/fold.ts and guarded by a test forbidding the pattern anywhere in src/ — which immediately found an eleventh site in adapters/repo-scene.ts that manual search had missed. Repro: tests/changeset.test.ts, 200 000 elements.

H3 — position claims finer than the frame's own error · TRUTH

date-fns aligns to a residual of 38.33 while MOTION_FLOOR is 3, so every position mark there was noise wearing a number: "moved 4.2 units" inside a frame uncertain by 38.

You cannot see motion smaller than your measurement error. Position marks now require moved > max(floor, frame.residual). Effect on date-fns: subjects offered for review fell 1605 → 303, with 1302 moving to display-only.

H4 — a ring I could see and could not reach · INTERACTION

The interface honestly said Review skips display-only subjects, and I immediately wanted to inspect one. Honest, and still a dead end. Added a second walk — "Inspecting display-only differences" — deliberately disjoint from Review rather than appended to it.

H9 — the build gate failed on an unreproducible measurement · INSTRUMENT

IV.2: composition stays linear failed the full suite repeatedly while passing every time it ran alone. node --test runs files concurrently, so contention can starve one scale and not the other and the ratio spikes for reasons unrelated to composition. Raising samples 5 → 9 reduced it and did not remove it.

Repaired by applying the project's own rule to its own instrument: a finding must reproduce before it is reported. An exceedance is now measured a second time and only a repeat fails. The threshold is unchanged, both observations are printed, and a real quadratic regression (2.78 against a 2.2 threshold) reproduces every time.

H5 — a large repository took 57 s to open · PERFORMANCE

Profiled rather than guessed. Two costs, both removed without changing a single output value:

phase	before	after	cause
import extraction	15.9 s	0.06 s	one git show per file; the question was always "which lines import relatively", never "what is in this file" — one git grep answers it
couplingMatrix	35.2 s	~1 s	Floyd–Warshall over a Map of Maps: two hash lookups in the innermost statement of a 4.3e9-iteration loop, now a flat Float64Array

repo	before	after
ky	539 ms	55 ms
zustand	499 ms	42 ms
zod	5 058 ms	293 ms
date-fns	57 s	5.4 s

Buildings, links, refusals and unplaced counts are identical on all six repositories — this bought speed and nothing else.

The density has a specific cause worth keeping: one date-fns commit touches 1606 files, and that single commit produces ~1.29 M co-change pairs on its own. The graph is dense because the history is.

Still open: n³ returns as a wall a few times larger, and the flat matrix is n² doubles — 21 MB at 1629 subjects, 800 MB at 10 000. Both ceilings are real and neither is guessed at.

H6 — zod reports 0 changes across 5 commits · RESOLVED, NOT A DEFECT

Confirmed by inspection: those six commits are chore:, docs: and ci: — they touch zero source files. Reporting no repository change was correct.

Recorded as resolved rather than deleted, because "the system said nothing changed and was right" is the outcome a suspicion is supposed to be able to reach.

H8 — an omission nobody could see · TRUTH

Recorded as "reported but unreachable". It was worse: unplaced never left acquisition. The API did not return it and no surface showed it, so a repository with 8 source files rendered 7 buildings and said nothing about the eighth. A reader counting buildings would conclude the eighth did not exist.

Refusing to fabricate a position is only half of it. The other half is saying that you refused.

The refusal now lives on the World — so it survives serialisation — carrying its basis and what is known:

lint-staged.config.ts — position is solved from co-change coupling, and this subject never changed in the same commit as another source file, so nothing licenses a position for it. known: repo/bytes=146 · repo/commits-touching=1

Measured cause, verified on three repositories: every unplaced subject has commits-touching = 1 and appears in no co-change edge. measured but unplaceable and unobserved are rendered differently, because a bare list would collapse them. An empty array is a claim that nothing was withheld.

Found on myself. The standard test fixture contains m1/tall.ts, added several sessions ago to hold the height reference steady. It appears in no commit, so it has never been placed — every measurement taken with that fixture ran against a world with a silently missing building, and nothing said so. The omission was invisible to the person who introduced it. Now asserted, so it cannot drift back into silence.

Recorded, not repaired

H7 — two counting units are correct and adjacent · AMBIGUITY

13 changed subjects and 25 recorded fact changes are different questions with different answers. Both are labelled and their relationship is stated, and a reader in a hurry may still read the second as a total of the first. No measurement supports a fix yet.

H10 — the silent fourth state · TRUTH

Found by applying the standing rule rather than by hitting it: a path must be placed, refused, or unobserved, and there must be no fourth. Checking whether a fourth existed found that it did, and it was the largest omission in the product.

Paths not matching the connector's extension set were dropped at the tree scan. Counted, never mentioned:

repo	in tree	entered the pipeline	dropped silently
ky	68	53	15
zod	583	406	177
date-fns	1903	1631	272
httpie (Python)	265	0	265

The last row is the failure at full strength. httpie rendered 0 buildings beside unplaced: [] — and by the semantics introduced hours earlier, an empty refusal list is a claim that nothing was withheld. Two true statements composing into a false one: this repository is empty. It has 133 Python files.

World.unobserved now carries the count, the vocabulary that was applied so the omission is checkable, and a breakdown by extension so a reader can see what kind of thing was skipped. Verified total on four repositories:

httpie    tree   265 = placed     0 + refused   0 + unobserved   265  ✓
ky        tree    68 = placed    53 + refused   0 + unobserved    15  ✓
zod       tree   583 = placed   406 + refused   0 + unobserved   177  ✓
ms        tree    23 = placed     7 + refused   1 + unobserved    15  ✓

H11 — a shallow clone fabricates a commit, and it moved the city · TRUTH

Found by the heuristic, not by an error: whenever you add a filter, ask where the excluded subjects go. Asking it of the history filter — rather than the path filter — found the worst defect of the session.

The boundary commit of a shallow clone has no parent, so git log --name-only reports it as touching every file in the tree. On httpie, 2db28ef "docs: fix typo (#1548)" appeared to touch all 265 files.

Those pairs become co-change couplings, and couplings solve position. The city's layout was partly derived from a commit that never happened that way.

	shallow	full history
ky commits-touching max	33	106
ky links	1006	673

More history produced fewer couplings — the inversion that gave it away.

Nothing can know what that commit really touched, so the file list is refused rather than guessed at, and the truncation is disclosed instead. The effect is large and it is the truth:

repo	before	after
zustand	49 buildings, 0 unplaced	5 buildings, 45 unplaced
p-queue	14 buildings	13 buildings, 1 unplaced

Zustand has 40 real commits and only 5 of its files genuinely co-change. The other 45 are now measured and honestly unplaceable, instead of positioned by a commit that did not touch them.

This invalidates earlier numbers in this file. Every fleet measurement above was taken on shallow clones, so link counts were inflated and commit counts were floors. The findings they supported still hold — they concerned attribution and scale, not absolute coupling — but the figures were measured through a window nobody had declared.

Method

1. Build a visible capability.
2. Use it on a repository the system has never seen.
3. Stop the moment something feels wrong.
4. Repair only what blocks truthful use; record the rest here.
5. Repeat.

Repositories exercised so far: ky, p-queue, zod, zustand, ms, date-fns.