fix(categorization): escape the NUL delimiter instead of embedding the byte

[KenTaniguchi-R]

The byte

engine.ts carried a literal 0x00 byte in its source, inside the template literal that builds the assignment-grouping key:

const key = `${assignment.categoryId}<0x00 byte>${assignment.source}`;

NUL is a sound delimiter here — it cannot occur in a category id or a source string, so the composite key cannot collide. The problem is writing it as a raw byte rather than the \0 escape: grep classifies the whole file as binary and silently drops matching lines. No error, no warning, just absent results.

$ grep -rn "categoryRules" src/ --include=*.ts
src/db/schema/categories.ts:43:export const categoryRules = pgTable(

One hit — reads as "nothing consumes this table."

$ grep -ran "categoryRules" src/ --include=*.ts     # -a: treat binary as text
src/lib/categorization/engine.ts:3:import { ... categoryRules ... }
src/lib/categorization/engine.ts:106:      id: categoryRules.id,
...
src/lib/categorization/engine.ts:112:    .from(categoryRules)
src/db/schema/categories.ts:43:export const categoryRules = pgTable(

engine.ts queries the table on lines 106–113 and runs the rules on every sync. The file was invisible to the search that would have shown that.

This is a live trap for anyone investigating this codebase — it produced a confidently wrong conclusion about the categorization pipeline while triaging #89.

The fix

\0 instead of the raw byte. Verified byte-identical at runtime:

escape === raw byte: true
codepoint at delimiter: 0
length: 10  expected: 10

file now reports ASCII text instead of data, so grep, git diff, and editors all read it normally.

Test

Adds the regression test that nothing covered: the grouping itself. Two transactions reaching two different categories through two different tiers (rule and merchant_default) in a single call must each land on the right category and the right source.

Verified the test has teeth rather than assuming it: collapsing the key to a constant (const key = "COLLIDE") fails this test — and only this test, out of the six in that file. The existing five pass either way, which is why this gap survived.

Full suite green: 753 passed / 113 files. Typecheck clean, npx eslint src tests clean.

Note: pnpm lint on this branch reports noise from .claude/worktrees — that is #117, not this change.

🤖 Generated with Claude Code

[KenTaniguchi-R]

Mutation gate: baseline comparison

	killed	timeout	survived	no-cov	errors	score
main's engine.ts	23	0	1	25	53	46.94%
this PR	20	3	1	25	53	46.94%

Identical score. The killed/timeout split differs (23/0 vs 20/3), but Stryker scores a timeout as killed, so both are 23 — the difference is runner load, not code.

Score-neutral is the expected result: the change is \0 for a raw NUL byte, verified byte-identical at runtime, so the mutants are the same ones; and the added test can only kill more, never fewer.

The file is structurally capped regardless — 25 no-coverage + 53 errors out of 102 mutants, because categorizeSyncedTransactions is DB-backed and vitest.stryker.config.ts excludes tests/integration/. That is #103.

Note the earlier typecheck · lint · test red on this PR was not code either — the self-hosted runner's pnpm store failed to reflink:

[ERR_PNPM_ENOENT] the source path is not an existing regular file, reflink
'/Users/taniguchiryusei/setup-pnpm/node_modules/.bin/store/v11/files/29/844c...'

Re-ran, and it passed.

Merging on the same basis as #102, #104 and #105.