fix(grep): correctly handle all flag shapes and never exceed raw output

[aeppling]

Summary

Make rtk grep produce correct output for every valid grep invocation and never cost more tokens than plain grep.

The #2333 arg-parsing rework routed many flag shapes through the grouping reparse, which silently dropped unparseable lines and returned success, producing wrong/empty results (and grep failed on --include) that pushed agents to unfiltered rtk proxy grep.

• Context (-A/-B/-C): grouped + compressed instead of dropped; header counts matches only.
• Unrepresentable output (-N, -I, --color, -p, --heading, --field-match-separator, binary notices): passed through verbatim, never a false "0 matches".
• Format/shape flags (-c/-l/-o/-Z/--column/-b/--vimgrep/--null-data): clean passthrough, no NUL leak.
• grep-only syntax (--include/--exclude/…): ripgrep fast path, falls back to system grep on rg error.
• Never-worse guard: never emits more than plain file:line:content grep.

Related

• Resolves grep: -r / -rn rejected on 0.42.4 (and -rln returns a false "0 matches" on latest develop) #2543 (-r/-rn rejected on 0.42.4; -rln false "0 matches" on develop).
• Supersedes fix(grep): strip ripgrep-only flags when falling back to system grep #2224 (rg-only flags leaking to the grep fallback) — folded into strip_rg_only + the rg-error fallback.

Test plan

• cargo fmt --all && cargo clippy --all-targets && cargo test --all — clean, 2242 tests pass.
• New integration tests: context compression, safety-net passthrough, NUL-free format/shape, grep-only-flag fallback, token savings, never-worse guard.
• Correctness battery vs bare grep across 18 shapes — no NUL, no false 0-matches, no dropped matches, never bigger than raw.
• Real hook path verified in tracking DB — routed (not proxy), 52% on bulky, 0% (never negative) on small.

[alexisLefebvre]

Nice — the unparsed_signal safety net looks like it'd catch this case by falling back to passthrough. One small thing I noticed: I couldn't spot a regression test for the exact #2543 shape, a bundled files-with-matches cluster (-rln/-ln). The new tests all use standalone flags (--column, -c, -o, -N, -I, …), so -l-inside-a-cluster isn't directly exercised. Might be worth a small test asserting rtk grep -rln <pat> <dir> lists the files rather than 0 matches, just to pin the original report — but no strong opinion if it's already covered elsewhere.

Repro + a 0.42.2 confirmation, in case useful: #2543 (comment)

🤖 Generated with Claude Code

[aeppling]

Thanks @alexisLefebvre , good catch. Added a regression test for the exact #2543 shape in 3a59d23: bundled_files_with_matches_cluster_lists_files asserts rtk grep -rln <pat> <dir> lists every matching file (and excludes non-matching ones) instead of a false 0 matches.

[EtaCassiopeia]

@aeppling Hey, just wondering if there was a reason my PR wasn't merged and the work was redone instead. Does the project accept external contributions?