fix(agent): keep the thought-level picker on a resumed conversation

[kaizhou-lab]

Symptom

Opening an existing claude conversation shows the model list but no
thought-level group. Sending a message does not bring it back. Leaving the
conversation and entering it again does.

Cause

A resumed conversation rebuilds its pickers from the persisted catalog while
the backend is still starting. That catalog stored only {id, label} per model:

{"available_models":[{"id":"opus[1m]","label":"Opus (1M context)"}, ...],
 "current_model_id":"opus[1m]"}

So resolve_current_model_efforts found nothing and get_config_options
omitted the reasoning_effort option entirely. Nothing re-publishes config
options when the live handshake lands, so it stayed missing for the whole
session. The second entry worked only because the task was warm by then and the
LIVE capabilities were used instead of the preload.

Both ends dropped the field:

• catalog_partial_from_caps projected id and label only;
• CatalogPreload::from_handshake set reasoning_efforts: Vec::new(),
  documented as intentional because the column did not carry them — a
  self-consistent pair that hid the gap.

Fix

Write side includes reasoning_efforts when a model has any. Read side pulls it
back out of the raw column: the shared ACP parser returns a SessionModelState
with no effort axis, so it cannot carry it through.

Models without an effort axis write no key at all. agy folds effort into the
model id (gemini-3.6-flash-high / -low) and codex has none; writing an empty
array for them would change the stored column for every backend in order to fix
one. Rows already in the database predate the field and load as "no efforts",
which is what they meant.

When it appeared

#609 (2026-07-23), which moved claude/codex onto the direct-CLI
SessionAgentTask and introduced CatalogPreload. Not related to the
Antigravity work (2026-07-31 onward) — that path projects no efforts at all, and
its catalog write-back only touches mode/model/slash-commands.

Tests

• efforts survive the write → read round trip, and feed
  resolve_current_model_efforts
• a model with no efforts writes no key (guards every other backend's column)
• a catalog written before the field still loads

Removing the write-side projection fails only the first, so the set neither
passes vacuously nor over-reaches.