feat: short-output auto-passthrough — skip compression when raw output is already minimal

[lg320531124]

Summary

Closes #2673

When command output is already minimal (≤5 lines, ≤500 bytes), skip the compression pipeline entirely and emit raw output unchanged. This prevents filter reshaping of short signal commands and avoids zero-savings compression overhead.

Problem

RTK compresses all command output. When output is already short, compression:

• Yields zero token savings (nothing to compress)
• Can reshape or inflate canonical markers (e.g. git push output, echo $?)
• Adds unnecessary latency (filter pipeline runs for no benefit)

PR #2667 adds hook-level passthrough for predefined signal commands, but:

1. Only covers the hook layer — rtk git status called directly still compresses
2. Allowlist-based — echo $?, which python3, hostname etc. are not in the list
3. Coarse heuristic — git status -uall (1000 lines) gets the same passthrough as clean git status (3 lines)

Solution

An output-length heuristic at the CLI execution layer that auto-passthroughs any short output, regardless of command name.

Both conditions must be met (AND logic):

• line_count ≤ 5 — covers git push (2 lines), echo $? (1 line), clean git status (3 lines)
• output.len() ≤ 500 — prevents passthrough for very long single-line output (e.g. minified JSON)

Interaction with #2667

Scenario	#2667 (hook)	This PR (CLI)	Result
git push (2 lines)	Hook skips rewrite	N/A (raw command runs)	✅ Raw output
rtk git status (3 lines)	N/A (already rtk)	Auto-passthrough	✅ Raw output
rtk git status -uall (1000 lines)	N/A	Exceeds threshold → compressed	✅ Compressed
echo $? (1 line)	Not in allowlist	Auto-passthrough	✅ Raw output
rtk git log -50 (200 lines)	N/A	Exceeds threshold → compressed	✅ Compressed

Changes

File	Change	Lines
src/core/config.rs	PassthroughConfig struct, effective_*() methods, config tests	+88
src/core/runner.rs	should_auto_passthrough(), integration in run_captured_filter(), tests	+110
docs/contributing/DESIGN-short-output-passthrough.md	Full design document	+393

Total: ~90 lines of production code (excluding design doc and tests). No new files, no new modules, no new dependencies.

Config

# ~/.config/rtk/config.toml
[passthrough]
short_line_threshold = 5    # ≤ 5 lines → passthrough (default: 5)
short_byte_threshold = 500  # ≤ 500 bytes → passthrough (default: 500)

• Default: Auto-passthrough ON (5 lines / 500 bytes)
• Opt-out: Set short_line_threshold = 1 + short_byte_threshold = 1

Test Plan

• Unit tests for should_auto_passthrough() (8 test cases)
• Config deserialization tests (4 test cases)
• cargo fmt --all && cargo clippy --all-targets && cargo test --all — all pass (2299 tests)
• Manual test: rtk git status (short → passthrough) vs rtk git log -50 (long → compressed)
• Performance: hyperfine comparison (passthrough should be faster for short commands)

Design Document

Full design doc: docs/contributing/DESIGN-short-output-passthrough.md

Covers: architecture, insertion point rationale, feature interaction matrix, edge cases, migration strategy, open questions.

[CLAassistant]

All committers have signed the CLA.

[lg320531124]

Friendly nudge — open ~28h, CI green. This is a perf win for the common case: when raw model output is already minimal (short responses), skipping compression avoids unnecessary token/latency overhead with no behavior change for longer outputs. +591 all new code + tests, no changes to existing paths. Happy to split into smaller commits or add benchmarks if that eases review.