fix: six correctness issues in author matching, excludes and lead time

[juangracia]

Six findings from an adversarial review pass. Each was reproduced before fixing and has a regression test.

HIGH

1. Author matching was case-sensitive, producing a silent zero

--fixed-strings does not fold case. A repo whose ident reads Alice@Corp.com:

$ gitrespect -a Alice@Corp.com    ->  14 added, 1 commit
$ gitrespect -a alice@corp.com    ->   0 added, 0 commits, exit 0, no stderr

Email domains are case-insensitive by spec and local parts are in practice, so people type their own address lowercase. The result was indistinguishable from "you did no work this period". --regexp-ignore-case now composes with --fixed-strings.

2. An uncompilable --exclude glob excluded nothing

filepath.Match reports a bad pattern as ErrBadPattern rather than a non-match, and all three call sites discarded the error:

$ gitrespect -e '[invalid'    ->  14 added, identical to no exclude, exit 0

Patterns are now validated once at flag-parse time and rejected loudly.

3. authoredLeadTime selected exactly the wrong commits

A committer date later than the author date identifies a rewritten commit, not a reviewed one. Cherry-picks, amends, rebases and whole-history rewrites all qualify, while commits that went straight to main have ct == at and were excluded. The sample was biased entirely toward rewritten history.

A repo with two healthy same-day commits plus one cherry-pick of two-year-old work:

Lead time (branch → main): Median 732.0 days (1 commit, authored → landed)

A repo that has run filter-repo rewrites the committer date on every commit, so every sample would be garbage.

Two guards: gaps longer than the analysed period are discarded, since work that took longer than the window did not flow through it, and a median now requires at least three samples. Verified both directions — a genuine rebase repo with four commits landing two days after authoring still reports Median 2.0 days (4 commits), while the cherry-pick case now reports no signal.

MEDIUM

4. dir/** excluded nothing while dir/* excluded everything

-e 'vendor/*'   -> 10 added  (correctly drops the whole subtree)
-e 'vendor/**'  -> 14 added  (drops nothing)

filepath.Match has no globstar, and the directory fast path that makes the single star recursive did not fire for **. Backwards from gitignore intuition: the more explicit pattern was the one that silently did nothing. ** now folds to *.

5. A leading ./ made a pattern a no-op

-e './vendor/*' excluded nothing. numstat paths are repo-relative with no ./, and the directory branch split on the first / giving . as the prefix, which never matched. Tab-completing a path is the obvious way to produce this. Patterns are now cleaned.

6. Non-ASCII filenames defeated every exclude

core.quotePath defaults on, so git wraps such paths in double quotes with octal escapes:

2	0	"caf\303\251.txt"
8	0	wk.txt

The surrounding quotes make every glob miss, so -e '*.txt' left 2 lines counted. All three numstat readers now go through a shared git.LogArgs that sets core.quotePath=false.

LOW

A daily average below 0.5 printed as a flat 0 lines/day because of %.0f. That is the headline number and "0" reads as "shipped nothing"; small rates now show a decimal.

Confirmed clean during the same pass

Worth recording since these were probed hard and held up:

• Week bucketing. (int(t.Weekday())+6)%7 is a correct Monday anchor; keys sort lexically and chronologically across a year boundary; no DST exposure because only calendar dates are used. Verified across Mon 2025-12-29 → Mon 2026-01-05.
• <> anchoring cannot be defeated by a crafted name. Committing as GIT_AUTHOR_NAME='Mallory <alice@corp.com>' makes git store Mallory alice@corp.com <mallory@evil.com>; the angle brackets are stripped from the name field, so the impersonation never reaches the searched string.
• Mailmap. --author matches the mailmap-resolved ident, so a .mailmap unifying two addresses is honoured. Desirable, but it means totals depend on each repo's .mailmap when -r aggregates across repos. Worth documenting later.

Verification

• go build, go vet, go test, gofmt -l clean
• 20 JSON paths parse with no NaN/Infinity; 7 HTML paths render with no unrendered directives
• Core fixture numbers unchanged across the whole scenario matrix
• Merge-commit lead time still reports via the merge path, unchanged

Follow-up in this PR: a Windows regression the CI caught

The first pass normalised patterns with filepath.Clean, whose separator is a backslash on Windows. That rewrote vendor/* to vendor\*, matching nothing, so --exclude stopped working entirely on that platform:

--- FAIL: TestAnalyzeExcludePatterns
    Added with exclude = 110, want 10

git reports forward-slash paths on every platform, so pattern handling now uses path rather than filepath throughout; filepath stays where it belongs, for real filesystem walks. Removing the shadow also surfaced that IsGitRepo and FindRepos took a parameter named path, now dir.

This is the cross-platform CI added in #4 paying for itself on its first real test.