# Path matcher maintenance — 2026-09-12

Contributor: ZJH-666-ZJH (朱鋆豪). Baseline: `156ba59`.

## Scope and design

The existing matcher allocated a full Boolean matrix both within a segment and
between path segments. Replace each matrix with one mutable row. Before
replacing a cell, save its old value for the next diagonal. A star consumes
from the new left-hand cell or matches empty from the old current cell.
Whole-segment `**` follows the same rule; ordinary segments use the character
matcher. Path validation, specificity and public API are unchanged.

## Reproducible quality evidence

`glob_wbtest.mbt` adds four test groups:

- 10,571 segment pairs: every pattern of length 0–4 over `a,b,*,?`, crossed
  with every value of length 0–4 over `a,b`. Independent recursive enumeration
  tries each possible wildcard span instead of building DP rows.
- 2,520 path pairs: patterns of 1–3 segments over `a,b,*,**`, crossed with
  values of 1–4 segments over `a,b`.
- Legal 1,024-character literal/wildcard inputs, 512-segment paths, and rejection
  beyond the length limit.
- Empty wildcard consumption, repeated `**`, Chinese BMP characters, and unsafe
  paths. This is not a claim of complete Unicode/grapheme conformance.

The new suite also passes against the baseline implementation, before restoring
the optimized implementation. The complete suite grows from 38 to 42 groups;
13,091 generated oracle pairs are assertions within two groups, not 13,091
separate test groups. Local wasm-gc, wasm and js strict checks/tests pass.
Native strict checking passes, but local execution is unavailable without a C
compiler. GitHub Actions additionally runs native tests and CLI smoke tests.
See `verification/` for the local logs; remote CI is a separate result.

## Quantitative bounds, not an RSS benchmark

For a pattern of P character units and value of V units, Boolean state slots
fall from `(P+1)*(V+1)` to `V+1`. At P=V=1024: **1,050,625 → 1,025 slots**,
a **99.9024%** reduction. For two 512-segment paths: **263,169 → 513 slots**,
a **99.8051%** reduction in the outer matcher table. Inner matching state is
additional. Neither figure measures process RSS, allocation bytes, throughput,
or end-to-end CI speed. Time remains O(P×V); no speedup is claimed.

Reproduce with `moon test --target wasm-gc --deny-warn` (or wasm/js/native),
`moon fmt --check` and `moon check --target wasm-gc --deny-warn`.
Public interface generation with `moon info` produces no interface diff.

## Position relative to CODEOWNERS

GitHub CODEOWNERS is the adjacent ownership-routing baseline, not a
performance benchmark. MoonChange combines path ownership with checks, labels,
line budgets, operation restrictions and release-note obligations in an offline
review plan. Its owner selection uses highest specificity, then last-rule tie
breaking: **do not treat it as a drop-in CODEOWNERS interpreter**. It neither
fetches GitHub evidence nor enforces branch protection. Input evidence must come
from a trusted integration. Binary rules require an explicit manifest because
the upstream diff metadata does not expose binary classification.

Reference for users comparing workflows:
https://docs.github.com/en/repositories/managing-your-repositorys-settings-and-features/customizing-your-repository/about-code-owners

## CI formatter compatibility

The first remote native job passed all 42 tests. The portable job exposed an
upstream formatter change (trailing commas in inline record expressions).
The maintenance branch applies those 11 formatting-only updates in seven files.
Local logs record the July toolchain before these formatter-only adjustments;
the remote CI format gate is authoritative for the September formatter.