chore(acp): refresh the bundled model catalogs

Regenerate both compiled-in snapshots from their sources: codex's from the
`@openai/codex` 0.153.4 nested under the pinned codex-acp 1.10.0 — the codex
codeg actually launches, not the one on PATH — and OpenCode's from a live
models.dev fetch put through `normalize_models_dev`, so the bundled bytes are
the shape the runtime cache writes rather than a hand-rolled approximation of
it.

codex gains `gpt-6-astra` and the hidden `gpt-daybreak-blue/red-latest`, and
drops nothing: retirement is still a flip to `hide`. Two shape changes travel
with it. `shell_type` is `unified_exec` everywhere it was `shell_command`, and
`supports_parallel_tool_calls` is gone from `ModelInfo` entirely — an ignored
unknown key now, the way `supports_reasoning_summaries` went before it. It stays
in `BOOL_FIELDS` because guarding it still costs nothing for anyone pinned to an
older codex. The strict-enum value sets and their nullability were re-probed
against 0.153.4 and are unchanged, so `enum_spec_for` only gains the version it
was last checked at.

`gpt-5.6-sol`'s priority moved 1 to 6, which makes `fallback_base_slug` — and
with it the default clone base for every custom entry and the default model
written as codex's root `model` — `gpt-6-astra`. The import test now takes its
clone source from `fallback_base_slug` rather than naming a slug, so the next
refresh cannot turn a priority reshuffle into a test failure.

OpenCode's catalog grows from 154 to 213 providers and 5418 to 7562 models.

`infer_context_window_max_tokens` extends its 258K lane to `gpt-6-*`, which fell
through to `None` until now and left those sessions with no context meter at
all. 258K is the generation's effective window as OpenAI's own catalog states
it — `context_window: 272000` at `effective_context_window_percent: 95` — and
gpt-6 carries the identical pair.