fix(cli): reject invalid --kinds values instead of silent fallback

[olympusbuildz]

Summary

buzz messages get --kinds "abc,9" silently dropped "abc" and returned kind-9 results with no warning. When all values were invalid, the parsed list was empty and the command fell through to the default kind set — the opposite of what the user asked for.

Root cause

cmd_get_messages parsed --kinds with filter_map(|s| s.trim().parse().ok()), discarding parse failures, and only overrode the default kinds when the resulting list was non-empty.

Fix

Replace the silent filter_map with a fallible parse that returns CliError::Usage on the first invalid value, naming the bad token. This removes the empty-list guard entirely — an all-invalid input now errors instead of reverting to defaults.

Why it matters

Silent data loss from malformed CLI input violates the agent-first contract: agents cannot inspect stderr interactively, so the CLI must fail loudly on bad input.

Test plan

cargo test -p buzz-cli --lib commands::messages::tests
# 34 passed (includes 2 new regression tests)

Risk / blast radius

CLI argument parsing only. No runtime behavior changes for valid --kinds inputs. Build-metadata only.

Closest work

none found

Fixes #6945

---

Reviewed by Hermes Agent · Olympusbuildz · 2026-08-28

[github-actions]

🔐 Codex Security Review

Status: review required for the current range.

The current range is 58cc4b7e9be71223f7898ea5f930a4248830ccfb...b1e87baa6f4dca690bdf6ac209243193c4976e9a.
A new review must complete for this exact range. When manual authorization
is required, a Block organization member must comment exactly
@buzz-security-review b1e87baa6f4dca690bdf6ac209243193c4976e9a to authorize a new review.
Any previous review applies only to its recorded range.

[fatima-n09]

@olympusbuildz Thanks for picking this up this matches the intended fail-closed behavior from #6945.
Invalid tokens should error (exit 1) instead of being dropped, and an all-invalid --kinds should not fall through to the default kind set. Happy to see the regression tests for both cases.