[oddharsh]

CLAUDE.md has carried two entries numbered 41 since 2026-08-24. #535's histogram re-encode entry landed on 2026-08-23, and #538 gave the cf-garage TypeScript-config entry the same number the next day. Twenty-four references pointed at gotcha 41 across twelve files: three meant cf-garage and twenty-one meant the histograms, and no reader could tell which without opening both entries and deciding from the sentence.

Pre-existing, unrelated to #783, which is where I found it.

The number moved, the position did not

The histograms entry is 46. Renumbering it to 42 would have collided, and shifting 42 through 45 up by one would have rewritten every cross-reference in a 6000-line file to fix a clash between two of them, over a file whose own layout note warns that a path rewrite cannot be run blind. So the list now reads 40, 41, 42, 43, 44, 45, 46-where-41-was. It was already out of order: 22 has sat above 21 since 2026-08-06.

The entry carries a paragraph explaining its own number, so a reader hitting 46 between 40 and 42 gets the reason without going to git.

The reference surface was wider than a grep of CLAUDE.md suggests

Eighteen of the twenty-four live outside it, in ci.yml, derivations.json, infra.json, a contract test, two tools/lib derive modules and five photo-pipeline files. AGENTS.md is a symlink to CLAUDE.md, so those are one file.

Two were buried mid-line where a fixed-width context grep missed them on the first pass, and both are real:

• config/derivations.json:303 — "the exact script whose standalone use caused gotcha 41"
• tools/photos/zenc-reproducible.ts:133 — "Suspect the INPUT bytes instead (CLAUDE.md gotcha 41)"

Every reference was resolved from its own sentence, never by pattern. No s/gotcha 41/gotcha 46/g was run anywhere.

One judgment call, flagged rather than buried

config/infra.json:329 read "Gotcha 41's lesson through another door" and fits both entries. The discriminator framing (a well-formed fake is the only probe that reaches the auth path) suits the histograms entry's self-agreeing check; the misread-refusal framing suits cf-garage's "which door it comes through". Git dates settle nothing, since both entries predate the line, and the commit that added it never names a gotcha.

I read it as cf-garage, on the refusal shape and the echoed word, and the reference now names cf-garage explicitly. Leaving it bare was not the neutral option: with 41 unique again, a bare 41 resolves silently either way, so the choice was going to be made regardless. Naming it makes the decision greppable and challengeable. Overrule it by editing three lines in infra.json.

The check

tools/contract-gotcha-numbers-resolve.test.mjs, two assertions:

1. Every gotcha number names exactly one entry.
2. Every gotcha N reference in the repository resolves to an entry, which catches a deleted entry rather than a duplicated one. That is the quieter failure: it reads as a pointer right up until somebody follows it.

Uniqueness only, never order. This entry and the 21/22 pair both break order, and asserting it would invite renumbering the file to satisfy a test, which is the change the check exists to make unnecessary.

Both assertions carry a floor, on the a-dict precedent: a scanner that has quietly stopped matching reports a clean pass over an empty list, and an empty list is trivially unique.

The heading scan is anchored at the section header deliberately. CLAUDE.md opens with several ordinary numbered lists at column 0 (the two homepage fragments, the converter's two rules, the three MCP deviations), and a whole-document scan reads those as gotchas 1, 2 and 3 and reports duplicates that are not there.

Three controls, each failing by name:

control	fails with
reintroduce the duplicate 41	gotcha numbers used more than once: 41
reference a gotcha 99	references to gotchas that do not exist: 99
move the section anchor	CLAUDE.md no longer carries the gotchas section header

Validation

• bun run lint clean, bun run typecheck clean (matches the baseline; all 328 files held by one of 10 programs), bun run test 681 pass / 0 fail.
• No served bytes move, measured rather than inferred. A build here and a build from a detached worktree at HEAD outside the checkout (gotcha 44) produce all 1829 files byte-for-byte identical.

Two things caught on the way: the type ratchet flagged a possible-null h.match(...)[1] in the new test, fixed rather than baselined; and my own comment carried the "not X, it is Y" negation, which a line wrap had hidden from the grep that should have caught it.

[github-actions]

Wire-size diff vs merge base

f1a62ca5 → 3803f3e5

Brotli q11, the quality the edge serves. Every number here is deterministic:
an unchanged file produces no row. Advisory — this check fails on nothing.

Client assets

No change. 25 files, 86.83 KiB Brotli total.

Pages

No change. 55 files, 578.04 KiB Brotli total.

Worker bundle

No change. 243.60 KiB gzip.

Dictionary transport

Family dictionary acquisition: no change, 14.41 KiB q11.

pd/: no change. 115 deltas, 472.30 KiB total.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
✅ Deployment successful! View logs	aadhar-sh	641d3af	Sep 10 2026, 07:06 PM