fix(desktop): surface model requests the harness rejected

[mikececco]

Problem

When a requested model has no byte-exact match in the harness catalog,
buzz-acp warns and starts the session on the agent's own default:

// crates/buzz-acp/src/pool.rs
"desired model {desired} not found in agent's available models — proceeding with agent default"

The only signal that reached Desktop was modelOverridden. That flag is false
in two different situations:

1. no model was requested at all, and
2. a model was requested and the harness rejected it.

reader.rs collapses both (if !model_overridden { … }), so the ModelPicker kept
showing the rejected pick as though it had taken effect. A wrong model was
indistinguishable from a right one, with nothing surfaced anywhere.

That matches the reports in #2692 ("picker persists but nothing applies it at
spawn", 3 independent repros), #4004 (composer-2.5[fast=false] runs as
fast=true), and #2265 (silent fallback when the default model needs credits).

Change

The harness now reports two additional fields in session_config_captured:

Field	Meaning
requestedModel	what was asked for, null when nothing was
modelApplied	whether the harness accepted it

The config surface derives unappliedModelRequest from the pair, and the
ModelPicker renders <model> not applied (with a tooltip naming the fallback)
when the live session is not honouring the pick.

Both fields are purely additive — model_overridden and all its existing
consumers are untouched. Harnesses predating the pair omit both, and a missing
requestedModel keeps those sessions out of the warning rather than
mislabelling them.

Why this PR also moves two blocks of code

desktop/scripts/check-file-sizes.mjs forbids any growth in files already
past the 1000-line cap, and both files this fix has to touch were over it. Per
the repo's own guidance ("split the file — never bump the limit"), each got one
mechanical extraction:

File	Before	Extracted to	After
commands/agent_config.rs	1112	config_bridge/session_payload.rs (ACP payload parsing, beside the types it produces)	989
shared/api/types.ts	1032	shared/api/configBridge.ts (config-bridge types, re-exported)	956

Both moves are code-identical — no behaviour change. Reviewing them as pure
moves should be quick; happy to split them into a separate prep PR if you'd
rather land them independently.

Tests

Three regression tests in config_bridge/reader_tests.rs pin the distinction
that was missing — including the case that makes this non-trivial: a session
running the harness default because nothing was requested must stay silent,
even though model_applied is false there too.

unapplied_model_request_is_surfaced_when_the_harness_rejects_it
unapplied_model_request_is_none_when_the_request_was_applied
unapplied_model_request_is_none_when_no_model_was_requested

Green locally:

• just desktop-check — pass (size ratchet included)
• just desktop-test — 3923 pass, 0 fail
• just desktop-tauri-test — 2109 pass, 0 fail
• just desktop-tauri-clippy — clean
• cargo test -p buzz-acp — 668 pass, 0 fail

UI note

The visible change is a single conditional label in ModelPicker.tsx, using the
same text-2xs text-warning treatment as the existing "restart to apply"
indicator next to it. I haven't attached a screenshot because the e2e mock
bridge doesn't populate unappliedModelRequest — I'd need to seed
buildMockConfigSurface to stage the state. Glad to add that plus a captured
screenshot if you'd like it before review.

Refs #2692, #4004, #2265