Fix pre-v2 CLI transcripts in spice search and spice trace output

[claudespice]

Summary

The Spice CLI was rewritten in Rust (spiceai/spiceai#9061, Jan 2026). Since then — and in every currently installable release, including stable v2.1.2 — spice search and spice trace render through TableOutput, which uses comfy-table's NOTHING preset: no borders, different column headers, OK/ERR statuses, and Time: <seconds> seconds. timing.

Three recipes still show the pre-rewrite Go CLI output, so a reader following them today sees something that looks nothing like their terminal.

Recipe	What the README showed	What v2.1.2 prints
docker	Rank 1, Score: 89.7, Datasets [spice.public.films] line output; Time: 32ms.	borderless Rank / Match / Score / Dataset table; score 0.8975; Time: 0.032 seconds.
azure_openai	bordered +---+ table, headers Key (path) / Dataset(s); Time: 650ms.	borderless table, headers Key / Dataset; Time: 0.650 seconds.
text-to-sql	TREE STATUS DURATION SPANID headers, check-mark status column	Tree / Status / Duration / Span ID headers, OK status

The docker score is the clearest break: the README printed 89.7 while the API response immediately below it — for the same match, same query — reports "_score": 0.8974827855112448. The old CLI displayed the score scaled ×100; the current one prints the raw score to four decimals.

Verified against

spiceai/spiceai at v2.1.2:

• bin/spice/src/output/table.rs — TableOutput loads presets::NOTHING (borderless; one space of padding per cell).
• bin/spice/src/commands/search.rs — headers are ["Rank", "Match", "Score", "Dataset"], or ["Rank", "Key", "Match", "Score", "Dataset"] when results carry a primary key; score is formatted {:.4}; the footer is Time: {:.3} seconds. N results.
• bin/spice/src/commands/trace.rs — headers are ["Tree", "Status", "Duration", "Span ID"] (+ Input/Output), status is OK/ERR, duration is {:>8.2}ms.

The azure_openai table keeps its Key column (its API response includes primary_key); the docker table has none (its API response has no primary_key).

Evidence

Static review — the recipes need Docker, Azure OpenAI, and OpenAI credentials, so they were not run.

The layout was reproduced from the renderer and checked byte-for-byte against a transcript already in this repo that was captured on v2.x — mcp/README.md shows real spice trace ai_chat output, and a re-render of those rows through the NOTHING preset rules matches it exactly. Every value in the corrected blocks is carried over from the transcript it replaces; only the layout, headers, status column, score scale, and timing format changed.

Two judgement calls worth a reviewer's eye:

1. docker is truncated to three results (...). The old transcript listed ten, but only the top three have exact scores documented — they appear in the recipe's own API response block. Rendering ranks 4–10 would have meant inventing four-decimal precision from one-decimal percentages, so the transcript stops at three and marks the truncation.
2. text-to-sql's trace is now flat, with no ├── glyphs. That is genuinely what the CLI prints today: in recurse_through_tree, new_indent returns String::new() whenever indent is empty, so the indent never grows past the root and the connector branch is unreachable. The mcp recipe's captured v2.x output is flat for the same reason. This looks like a CLI regression rather than intended behaviour — happy to revert this hunk if the tree rendering is going to be restored upstream instead.