feat: tool card rendering polish — monospace titles, raw-input preference, gemini noise filter

[kevin-david]

Why

Two tool-card rendering issues, both surfaced while testing the adapter against gemini-acp but applicable to any agent that emits paths/commands in tool calls.

What changes

1. Render code-y titles as monospace instead of **bold**. Execute/read/edit/write/delete/search-kind titles are user content — shell commands, file paths, search patterns. Inline code reads better than bold and avoids accidental markdown collisions in command strings. Descriptive labels (Update Topic Context, Invoke Subagent) stay bold via the CODEY_KINDS allow-list.

2. Restructure buildTitle to prefer rawInput over displayTitle. Some agents (gemini-acp) emit prettified displayTitle strings with capitalized first letters that don't match the actual paths on disk (Life/... for the real life/...). Prefer the raw command, file_path, pattern, etc. so the title reflects truth.

3. Drop the final capitalize(entry.name) fallback. Gemini puts the real content directly in entry.name with no rawInput/displayTitle, and capitalizing the first letter corrupted real paths. Argless commands like Pwd are still lowercased via the EXECUTE_KINDS short-identifier regex branch above.

4. Extend evaluateNoise to flag gemini's internal bookkeeping tools (Update Topic Context, Invoke Subagent) as noise so they're hidden at low/medium and only shown with 👁️ at high.

Test plan

• Verified life/finance/foo.md no longer renders as Life/finance/foo.md
• Verified pwd (was Pwd) renders correctly via EXECUTE_KINDS branch
• Verified Update Topic Context no longer appears in medium-mode tool cards
• Reviewer: test on Claude Code too to confirm no regression for agents that provide proper rawInput

🤖 Generated with Claude Code

[0xmrpeter]

The four issues you're tackling here are all real and worth fixing — especially the gemini re-emission deduplication, which would have caused very noisy Discord threads. Good investigation work!

A few things to resolve before this leaves DRAFT:

Blocking:

1. onToolCall dedup guard calls upsert() instead of get() — upsert overwrites the existing entry with a fresh "running" state, discarding any completed/failed status that prior onToolUpdate calls may have already applied. The guard should read the existing entry (.get()) and re-render it as-is, not replace it.

2. Existing test regression — formatting.test.ts:82 asserts **config.ts** (bold) for a read kind spec, but this PR adds read to CODEY_KINDS which makes it render as monospace (`config.ts` instead). The test needs updating to match the new expected format, otherwise CI will fail.

3. Dead code after restructure — the old EXECUTE_KINDS.has(kind) block (now containing only the description check) is unreachable for any tool that has a command or short-identifier name. It should be folded into the new top block or removed to avoid confusing future readers.

Non-blocking:

• includes("topic") in evaluateNoise is a broad substring match — it'd silently suppress any future tool whose name contains "topic" (e.g., get_forum_topic, create_topic). An exact-match Set of known gemini internal tool names would be safer and more explicit.
• isCodeyKind is a one-liner wrapper around CODEY_KINDS.has() with no added readability — could be inlined at the call site.

Looking forward to seeing this land once those are sorted!