feat(tui): show the model that actually served a routed turn - #12

The turn footer rendered only the CONFIGURED model name, so every turn on
a router route read "Build · Firerouter glm" no matter which member model
answered. The route slug is not a model, so that line was hiding both what
ran and what the turn cost.

Renders the configured name decorated with what served it:

  Build · Firerouter glm (glm-5p3-flash)
  Build · Firerouter glm (glm-5p3-flash → glm-5p3)    multi-step turn

A direct provider echoes its own id back on every response, so a served
list that resolves to the configured name is dropped and every non-routed
footer stays byte-identical.

Covers all three places that render the line - the interactive TUI, the
markdown transcript export, and `opencode run` (both session replay and
the live turn). The live path never saw the assistant message, so the
served ids are carried to the footer as new FooterState, reset at
turn.send so one turn's route cannot decorate the next turn's summary.

Also regenerates the SDK types and openapi spec, which #6 left stale, and
fixes the repo's one lint error (a Tailwind CSS escape read as a
deprecated octal escape) so `bun run lint` is clean.

Claude-Session: https://claude.ai/code/session_017iG7xA9wRvNyFYZSXm8inU

The live turn summary took its model from the COMPOSER SELECTION, so
switching models while a turn was in flight labelled the finished turn
with the model that had just been selected - naming a model that never
saw the turn, with no sign anything was wrong.

The assistant message records the identity the turn was actually
dispatched to, so carry that to the footer alongside the served ids and
prefer it. The selection survives only as the fallback for a turn that
produced no assistant message at all.

FooterState.served becomes FooterState.turnModel, one field carrying the
identity and its served ids from one event with one reset, and patch()
now merges it on key PRESENCE so turn.send's explicit undefined actually
clears it instead of being read as "unchanged".

Precedence lives in turnSummaryModel() rather than inline in the event
handler, so the rule this commit exists for is directly testable; the
test fails if the two branches are swapped back. Drops appendServedLabel,
which is unreachable now that served ids never arrive without an identity.

Claude-Session: https://claude.ai/code/session_017iG7xA9wRvNyFYZSXm8inU

…lection

Review round 1 (codex) found three real defects in the previous two commits.

1. A prompt produces one assistant message PER STEP - session/prompt.ts
   loops and creates a new Assistant each pass - while the run footer
   prints ONE summary for the whole prompt. Keeping the last message's
   record meant a tool-calling turn that routed A then B reported only
   B, silently dropping exactly the multi-model sequence this feature
   exists to show. Served ids now accumulate in first-seen order across
   every assistant message sharing a parent user message, keyed by that
   parentID so a new prompt starts a new record on its own evidence.

2. A turn that failed before recording any model still fell back to the
   composer selection, which is mutable while a turn is in flight - so a
   failed turn could be confidently labelled with a model chosen after
   it died, reproducing the bug the previous commit set out to fix.
   turn.send now snapshots the dispatched identity, which is the last
   moment the selection IS the truth, and the summary says "unknown
   model" rather than borrowing a selection.

3. Suppression compared resolved DISPLAY NAMES, so two different ids
   sharing a friendly name hid the served id - different versions,
   routes and prices read as identical. It compares raw ids again; only
   an id identical to the requested one is redundant.

Accumulation lives in session-data rather than the footer's patch merge:
turn.send writes a seed, so a merging patch would have folded the
previous turn's models into it.

Also pins the turn.send reset as a present-key-carrying-undefined (a
value-based merge would drop it), and pins subagent isolation, which a
reviewer suspected was broken and which nothing asserted.

Truncation of the summary line at narrow widths is filed as #13.

Claude-Session: https://claude.ai/code/session_017iG7xA9wRvNyFYZSXm8inU