fix(pytest): pass --version through unfiltered

[maxmilian]

Summary

• rtk pytest --version returned Pytest: No tests collected instead of pytest's version string. Fixes rtk pytest --version returns 'no tests collected' — wrapper consumes the flag #2096.
• Route --version around the output filter and forward pytest's output verbatim.

Root cause

pytest --version prints its banner and exits without running a test session, so filter_pytest_output finds no passed/failed/skipped summary line. build_pytest_summary then hits the all-zero-counts branch (src/cmds/python/pytest_cmd.rs:183) and returns the hardcoded "Pytest: No tests collected" — the real banner is discarded. (The reporter's "wrapper consumes the flag" hypothesis is close; the flag is forwarded, but the filter swallows the response.)

Fix

In run(), detect --version before flag injection and forward the command with an identity filter (no -q/--tb=short/-rxX injection, no compression), so pytest's output reaches the caller byte-for-byte. Honors RTK design principle #1 (Correctness over Token Savings): a flag that explicitly asks for informational output must not be reformatted into a lossy summary.

Scope (deliberately narrow)

• --version only. -V/-VV are version-dependent — pytest 9's bare -V runs a full session, so treating it as informational would dump a raw, unfiltered test run (a regression). Encoding that fragile per-version behavior isn't worth it; --version is the reported, unambiguous case.
• --help/-h are out of scope: clap intercepts them at the subcommand layer and prints rtk's own wrapper help — they never reach run().
• --collect-only/--co left untouched — already addressed by fix(pytest): handle --collect-only output correctly #925.

Tests

• test_version_banner_swallowed_by_filter — documents the buggy filter behavior (banner → "No tests collected").
• test_is_informational_flag — --version matches; -V, -VV, --help, -h, -v, --verbose, --collect-only, and test paths do not.
• Full suite: 1948 passed, clippy clean, cargo fmt --check clean.

Manual verification

$ rtk pytest --version          # before: "Pytest: No tests collected"
pytest 9.0.3                    # after: real version
$ rtk pytest -V                 # still filtered (no regression)

[maxmilian]

Note for the PR Testing Checklist: the "token savings ≥60%" item is N/A here. This is a correctness fix, not a new/compressing filter — rtk pytest --version was being collapsed to "No tests collected" because the filter has no test-session summary to parse. The fix bypasses the filter for --version so pytest's real output reaches the caller. This applies design principle Correctness vs Token Savings: a flag explicitly requesting informational output must not be reformatted into a lossy summary.

Remaining checklist items:

• Unit tests added (test_version_banner_swallowed_by_filter, test_is_informational_flag, combined-args detection)
• cargo fmt --all --check && cargo clippy --all-targets && cargo test — all green (1948 passed)
• Manual rtk pytest --version → real version string; rtk pytest -V confirmed to stay on the filtered path (no regression)

Scope is deliberately narrow (--version only) for correctness — -V/-VV are pytest-version-dependent and --help/-h are handled by clap before reaching the wrapper; details in the PR body.