[decision_models] Compact the browser front's state for Laya's window

[chaimaerachdi]

Why

The browser front sends every decision model the same state it sends Jev: a JSON object per element row, the
full page text, and ten actions of history. docs/benchmarks.md shows that state runs 18,785-23,654 tokens
on the Allrecipes run. Laya's window is 512 to 1024 tokens. The state fills the window before a single
instruction token is spent, which is why --model laya raises MODEL_SERVICE_CONFIG_ERROR on most real
pages today.

What

• laya_state() in s1a/decision_models/laya.py, called from LayaModel._decide before every request:
  • page.text dropped. It's for the chat model's DONE answer, which Laya never writes; the choice heads
    already carry each candidate's own text.
  • Each element row: one short line (idx role label =value [flags]) instead of a JSON object with
    repeated key names.
  • recent_actions: last 3 kept instead of 10.
  • Anything that isn't this shape (tool front, rails, a plain string) passes through unchanged — those
    already fit the window.
• On by default; LAYA_COMPACT_BROWSER_STATE=0 turns it off.
• docs/decision-models.md, CHANGELOG.md updated.

Roughly a tenfold reduction in the JSON-shaped state's size on the fixtures used in the tests below.

What this does not do

This closes the "state is bigger than the window" gap, not the "is a filled window, on the real 0.4B
checkpoint, actually under 7s end to end on a real page" question. That needs uv sync --extra laya, the
real convaiinnovations/laya weights, and a live browser run — I don't have a GPU or a live browser session
in the environment I built this in, so I have not run that. Exact repro:

```
uv sync --extra laya
LAYA_MAX_LEN=1024 uv run s1a run flights --model laya
```

If 1024 still isn't enough on a heavier page, LAYA_MAX_LEN needs to go up further (2048, 4096); that knob
is unchanged by this PR.

Verification

• uv run ruff format --check s1a/decision_models/laya.py tests/test_decision_models_laya.py
• uv run ruff check s1a/decision_models/laya.py tests/test_decision_models_laya.py
• uv run pytest tests/test_decision_models_laya.py -q: 34 passed, 2 skipped, 6 subtests passed
  (FakeLayaAgent, no torch/weights needed)
• uv run pytest tests/ -q --ignore=tests/test_browser_policy.py: same 19 pre-existing failures on
  main before and after this change (missing ty binary and other environment-only gaps in the
  sandbox I built this in; tests/test_browser_policy.py fails to collect on main too, before this
  branch — openjiuwen.harness.schema.decision_policy isn't in the pinned [email protected] tag
  installed there). Neither is touched by this change.
• Real-model / real-page latency and accuracy numbers against the <7s goal — not run, see above.