[kvlonge]

Why

Every sqb invocation except --version and warm compile replays started Python, even help, usage errors and commands that need no Python. Starting it costs about 0.1 s before any output.

Changes

• sqb parses arguments natively and prints help, version and usage errors itself, with Python's int(), terminal width and colour rules. A --vars value or a non-text argument still goes to Python.
• clean, a bare sqb, skills, init and rules list|show run in sqb. The decision is made before any side effect, from the arguments and cheap project facts. Python runs instead for --debug, sinks or event exporters, a project file Python would reject, custom rules, rule options, PYTHONPATH/PYTHONHOME, or when the installed package cannot be found. A failure before the command writes anything also reruns in Python with the same arguments, so Python reports it.
• Observed native commands write the same compute log capture and skill freshness notice as Python. Ctrl+C ends sqb by SIGINT, as it ends Python; a closed stdout gives Python's status, message and capture.
• Project scaffolding and the rule listing text are shared by both paths. The Python CLI entry point still serves every command, so its handlers stay as thin adapters.
• Recall does not record native commands, the same policy as launcher replays. The docs say so.

Verification

• Parity E2E over 27 argument lines (help, usage errors, version, clean, skills, init, rules): stdout, stderr, exit code and compute log files match the Python CLI byte for byte, and Python never starts. Further tests cover fallbacks to Python, Ctrl+C and stdout without a reader.
• Rust unit tests; init, skills, rules and Recall E2E/integration suites; clippy -D warnings, fmt, fensu, ruff, ty, boundary, layering and docs-skill checks.
• Same machine, 5 runs each, median instructions:u / wall: --help 381M/97 ms → 14.6M/11 ms; usage error 381M/96 ms → 13.8M/12 ms; rules list 566M/177 ms → 21.8M/43 ms.

[kvlonge]

Closing in favour of one combined PR (H2 + fensu 0.27 + SharedMergeTree on top of the merged deadlock fix) so these share a single CI run; the branch is being fixed for an interaction between H2's broken-pipe parity tests and the merged changes.