feat(grep): fold shared path prefix in file-list output

[israellot]

Problem

rtk grep -l / -L / --files route through passthrough() and print paths verbatim — one per line, each repeating the same long directory prefix. When an agent passes an absolute search root (e.g. grep -ril foo /home/user/project/...), every line restates /home/user/project/.... That shared prefix is the single most strippable part of the output, yet it was forwarded uncompressed and track_passthrough recorded 0 tokens saved for the command.

$ rtk grep -ril websocket /home/u/proj/findings /home/u/proj/methodology
/home/u/proj/findings/vpn_login_websocket_hang.md
/home/u/proj/findings/meet_realtime_listen_detect.md
/home/u/proj/methodology/client_server_time_sync.md
...                       # ← prefix repeated on every line, 0% savings

Change

Add fold_path_prefix (in grep_cmd.rs): factor the longest common directory prefix into a one-line header, then group the remaining tails by sub-directory — the same dir/ file1 file2 shape find_cmd already emits, so no new output format is invented (Transparency principle), and the path-grouping logic mirrors existing code (Extensibility principle).

/home/u/proj/ (3 files)
findings/ vpn_login_websocket_hang.md meet_realtime_listen_detect.md
methodology/ client_server_time_sync.md

Wired into passthrough(), which is where the raw-path leak happened (the -l/-L/--files format-flag branch and the unparseable-shape fallback both funnel through it).

Correctness guarantees

• Lossless — every input path is recoverable as header_prefix + dir + "/" + file. No truncation, so no CAP_* / recovery hint needed.
• Never-worse guarded — wrapped in guard::never_worse; non-path passthrough output finds no common prefix (None) and is emitted verbatim.
• Component-boundary safe — the common prefix is computed on whole path components, so /a/foobar and /a/foobaz share /a/, never /a/fooba.
• Returns None (caller keeps raw) for <2 paths, no shared directory, bare filenames, or a degenerate /-only prefix.
• Tracking switched from track_passthrough (0/0) to track so the savings are actually measured. When nothing folds, in == out — no dilution.

The default grouped-match grep path (file:line:content) is unchanged.

Savings

Honest numbers: pure file-lists are bounded by the filenames themselves (irreducible signal), so this lands ~30–55% depending on prefix depth and file count rather than the 60%+ typical of verbose-output filters — but it converts a measured 0% → positive on a common agent operation. Examples measured locally:

• Shallow 2-dir case (-ril, 5 files): ~50%
• Deep Unity-style abs paths (3 files): 38% (252 → 156 bytes)

Testing

• 9 new unit tests for fold_path_prefix (absolute, relative, sub-dir grouping, component-boundary, single-path/no-common/bare-filename/root-only → None, size-shrink invariant).
• cargo fmt --all --check && cargo clippy --all-targets && cargo test all pass (2268 passed, 8 ignored).
• grep smoke tests pass; the unrelated smoke failures (curl/wget/pnpm/rewrite/init) are pre-existing and environment-bound (verified identical with this change stashed).
• Manual before/after verified on the motivating grep -ril case.

Docs

Updated src/cmds/system/README.md — the line that said -l/-L "bypass RTK filtering and run raw" now notes the path-prefix folding.

[CLAassistant]

All committers have signed the CLA.

[KuSh]

Thank you for this. The file this changes, src/cmds/system/grep_cmd.rs, was removed on develop by #2641 (merged 2026-06-26), which rebuilt grep and rg around a shared Engine model in src/cmds/system/search.rs, and its passthrough path also gained a streaming mode that never buffers the whole output. That means the fold_path_prefix helper and its tests no longer have a place to attach, so I am closing this. The underlying gap is real: file-list passthrough still prints each path in full and records no savings, so an issue or a small PR against search.rs passthrough would be very welcome. If I have misread part of this, please comment or reopen.

[israellot]

Thanks @KuSh, that reading is correct. Re-targeted at the search.rs passthrough in #4170: same fold, now applied to grep -l/-L, rg -l, and rg --files under never_worse, with the streaming stdin path left untouched.