feat(tools): use --format llm for the body, keep each tool's envelope

[KageBinary]

The middle path

Nine tools fetched --format json, parsed it, and hand-rendered markdown. The rendering carries real value — a header the model orients on, an empty-graph message, an error envelope — but the body is usually a table or list that --format llm already emits, 2-4x smaller than the JSON it was rebuilt from.

So: keep the envelope, swap the body. A tool that gets llm text back emits its own header and then the records verbatim; a tool that gets null runs its existing JSON path untouched.

Tools wired: stats subsystems smells trace locate rank inventory explain neighbors.

Backwards compatibility

The floor is per command. --format llm arrived tier by tier — Tier 1-4 in v0.7.0, Tier 5 (explain, read) in v0.9.2 — so ix-explain carries a different floor from the other eight. Confirmed against the tags containing each tier's commit.

Getting it wrong would not have raised. ix does not validate --format. Every renderer is if json … else if llm … else text, so an unrecognised value falls through to human text and exits 0. On 0.9.1, explain --format llm returns prose, successfully. An ungated call would have handed the model prose dressed as records, with nothing to catch.

The same property is what makes this safe to ship: there is no version of ix on which asking for llm breaks. The floors buy output quality, not crash-avoidance — which means the downside of a mistake here is degraded output, never a broken tool.

Pro commands are excluded outright. briefing and decisions come from @ix/pro, which declares only text|json; there is no llm renderer at any version, so no gate can help. They are absent from the table, and a test walks every tryLlm call site to prove none of them names one.

Where the envelope needed more than a header

• ix-stats defers to JSON when the records report an empty graph, so "run ix map to index the codebase" survives. That is the most useful thing this tool says, and it is worth one extra call in the rare case. A substring check, not a parse — the fast-path never interprets its output.
• ix-neighbors stacks several sections under one header, so the fast-path keeps the per-section label. Unlabelled records would leave the model unable to tell callers from callees.

Carried across from ix-cursor-plugin's implementation

• error code= deferral. ix reports some failures as a record on stdout with exit 0; checking only the exit status would forward an error line as a result.
• IX_DISABLE_LLM_FORMAT=1 kill switch.
• Version probe memoised per process; fails closed.
• llm output is never parsed, only forwarded — and run through redactSecrets first. (The JSON path does not scrub today; bringing it up to match is a separate change.)

Tests

24 new in tests/llm.test.ts: the table's contents for all three categories, gte across each position, diff --content staying on text, the error code= detector including cases that must not fire, the kill switch, and a wiring check that every fast path keeps its header.

bun test → 112 pass. The 3 PluginHookContract failures are pre-existing (@opencode-ai/plugin is not installed in this checkout) and unchanged — verified against a clean tree before starting.