refactor(git): tokenize git show's objects instead of walking clusters

[KuSh]

What the TODOs were

src/cmds/git/git.rs carried two TODO(after #3681) markers left by the author of the
blob-show path (#3265), both waiting on the ValueSpec factorization that landed with #3681.

1. flag_token_consumes_next — the short-cluster walk. show_positionals found a
git show's object arguments with a hand-rolled scan: position(|a| a == "--") for the
boundary, starts_with('-') for flags, and a per-char walk of each short cluster that
re-queried the flag table by rebuilding format!("-{c}") for every character. The semantics
it encoded — the first value-taking short flag in a cluster takes the remainder as an inline
value, or the next argument when it is the cluster's last char — is exactly Attachment plus
the solo_only rule the tokenizer already models. It is now:

let tokens = tokenize_git_diff_args(args);
arg_tokenizer::before_dashdash(&tokens)
    .iter()
    .filter(|t| t.is_free_positional())
    .map(|t| t.text)
    .collect()

Three functions and a per-char allocation go away.

One grammar per subcommand. This reads git show under diff_takes_value, not
log_takes_value, matching commit_or_stat_route's existing tokenize_git_diff_args for the
same subcommand. Measured against real git, not assumed:

invocation	real git
git show -wl 100 HEAD	ok — -l is the rename limit and clusters
git show -wl HEAD	error: switch `l' expects an integer value
git show -n 1 HEAD	ok
git show -wn 1 HEAD / -pn 1 HEAD	fatal: ambiguous argument '1'

So -l is value() and -n is solo_only() — the diff table. Under log's table -l is
solo-only, which would leave the 100 looking like a second object and silently drop the
blob window.

2. probe_is_blob — dropped, deliberately. The suggestion was a flag pre-filter to avoid
the cat-file subprocess on the common path. There is nothing conservative left to skip: the
probe only runs once a free positional already looks like rev:path, and after this change
that set is strictly smaller than before. An ordinary commit show never reaches it, and on the
blob path rtk git show HEAD:file.txt measures 3.7 ms ± 0.4 ms end to end — probe included,
well inside the 10 ms budget and unchanged from the merge-base. Any pre-filter would trade a
non-problem for a risk to the one decision only git can make, so the marker is removed rather
than acted on.

Also fixes the doc link on log_wants_raw_shape, which pointed at requests_raw_log_output —
a #[cfg(test)] helper rustdoc cannot resolve outside a test build. Other broken links in the
tree are pre-existing and left alone.

Differential verification

A binary built from the merge-base and one from this branch, run over a matrix of git show
invocations in a fixture repo under LC_ALL=C, with stdout, stderr and exit code compared
byte for byte.

• Matrix 1 (55 cases): plain commit show, rev:path blob show, :0:path index blob,
  -wG a:b HEAD:blob, -pS a:b, -S 'url:1' HEAD, -L 1,2:file, -- <pathspec>,
  --textconv, --filters, --ext-diff, --no-textconv, the magic pathspecs
  :(exclude) / :! / :^ / :/text, :: / :, --stat / --numstat / --name-only /
  --pretty / --format, --word-diff / --color-words / --diff-merges, -s,
  --grep -p, -20, bare -, and multi-object shows.
• Matrix 2 (39 cases): every short flag whose log and diff spellings disagree, each carrying a
  colon operand next to a real blob so grammar drift shows up as a routing change —
  -wl a:b, -pn a:b, -cl a:b, -wG/-wS/-wI/-wO/-wL, -U 3 vs -U3, -M50,
  -C/-B, -Gfoo, -wGa:b, -lw 100, -wnl 1, a -- as a flag's would-be value, and
  -- combined with each.

Result: 94/94 identical, 0 differences. The runner was sanity-checked against itself first
(same binary twice) to confirm it is deterministic.

tests/git_show_blob_differential_test.rs, the seeded 600-iteration fuzzer that already guards
this path against git cat-file -t ground truth, also stays green.

Regression tests

Behavioural, CI-run (not #[ignore]), against the real binary and real git, in
tests/guard_integration_test.rs:

• git_show_cluster_flag_value_is_not_mistaken_for_the_blob_object — -wG a:b HEAD:big.txt
  must still window the blob. Red when show_positionals is reduced to a
  starts_with('-') scan.
• git_show_rename_limit_clusters_under_diffs_grammar_not_logs — -wl 100 HEAD:big.txt must
  still window the blob. Red under log_takes_value, green under diff_takes_value. This
  is the test that pins the one-grammar-per-subcommand choice.
• git_show_blob_spec_after_double_dash_is_a_pathspec_not_an_object — a blob spec past --
  must print the commit, not the file. This one is a pin rather than a fail-before test: the
  cat-file probe and the can_window gate already make it unobservable on their own, which
  is the defence-in-depth working as designed.

Both fail-before results were confirmed by temporarily reverting, running red, and restoring.

Performance

hyperfine --warmup 5 -N, merge-base vs this branch:

invocation	merge-base	this branch
git show HEAD	6.6 ms ± 0.9	6.7 ms ± 0.5
git show HEAD:file.txt	3.8 ms ± 0.3	3.7 ms ± 0.4
git show -wG a:b HEAD:file.txt	3.6 ms ± 0.4	3.7 ms ± 0.4

Within noise, inside the <10 ms budget.

Out of scope

The output-shape routing refactor proposed in #3910 — log_wants_raw_shape,
diff_wants_raw_shape, show_wants_raw_shape and body_is_suppressed are untouched here so
that work can land cleanly.

Related issue

Refs #3954 — that issue targets flag_token_consumes_next's hand-maintained value-taking table,
which this PR deletes in favour of the shared diff_takes_value grammar, so the two-tables-drift
root cause it describes is closed here. Not Fixes, because its specific claims no longer hold:
its one reproducing example (--ignore-matching-lines a:b HEAD:big.txt) already windows on
develop (36090 B raw -> 8243 B), and git 2.53.0 rejects the three flags it names in separated
form (git show --notes x HEAD:big.txt -> fatal: ambiguous argument 'x'; same for
--conflict-marker-size and --submodule), so adding them to any table would make RTK consume
a token git treats as a revision.

[rtk-wshm-sync-bot]

wshm · Automated triage by AI

📊 Automated PR Analysis

♻️ Type	refactor
🟡 Risk	medium

Summary

Refactors git show's argument scanning to reuse the arg_tokenizer's Attachment/ValueSpec grammar instead of a hand-rolled short-cluster walk, resolving two TODOs left pending on the prior ValueSpec factorization. It removes the now-unnecessary probe_is_blob pre-filter TODO deliberately, and adds differential and regression testing showing behavior parity with real git.

Review Checklist

• Tests present
• Breaking change
• Docs updated

---

Analyzed automatically by wshm · This is an automated analysis, not a human review.

[aeppling]

lgtm