feat(export): subagentType and model on per-record exports; --by-agent follow-ups

[ozymandiashh]

Follow-ups to #710, as offered in its review. Three pieces, all small:

1. Export (issue Feature request: agent × model cross-breakdown (e.g. codeburn agents --by-model) #708's second ask, deferred there): new records.csv with one row per priced call including subagentType and model; sessions.csv gains both fields appended after all legacy columns, so by-index CSV consumers keep their positions; JSON fields are additive; escCsv formula-injection escaping applies to the new values.

2. Collision-safe main bucket: agentType is free-form text from the transcript meta, so the bare main sentinel merges a real subagent literally named main into the main-session bucket. The sentinel becomes (main); a test pins that an agent named main stays distinct across table, markdown, json and csv.

3. README: one sentence noting the inherited --min-cost 0.01 default can silently hide sub-cent agent buckets (common at agent granularity) and that --min-cost 0 shows the full inventory.

[iamtoruk]

Verified locally: tsc clean, full suite green apart from the #681 electron failures, and both behaviors mutation-pinned; reverting the (main) sentinel to bare main fails exactly the collision test, and blanking subagentType in the record rows fails exactly the two export pins. Ran a real export over yesterday's data: records.csv carries subagentType and model on 631 real agent rows, and sessions.csv appends the new fields after every legacy column so positional consumers keep working, exactly as described. The (main) collision catch on #710 was a good one, thanks. This also closes out the export follow-up promised on #708.