docs(mcp): laya_decide row must list min_confidence; every tool row must name every handler param

[aashish254]

What

docs/cli-mcp.md's laya_decide row listed `state`, `schema`, optional `model` -- but
laya/mcp/server.py:501 registers
laya_decide_tool(state, schema, model="auto", min_confidence=None), and the tool's own
description spells out min_confidence ("a field whose answer falls below it comes back as
null in values"). The row now names min_confidence too.

Adds test_cli_mcp_page_tool_argument_rows in tests/test_mcp.py. For each of the six
single-shot tools on the docs page (status, route, predict, shortlist, preset,
decide), the gate reads the third |-delimited table cell and asserts every public parameter
inspect.signature sees on the registered handler appears backticked in the cell. The
pre-fix laya_decide cell is banned by direct substring so the check fails if the row reverts.

Why

The MCP docs page is the fourth copy of the tool's parameter set, after the validator error
message, each tool's registered description, and the README. test_batch_item_shape_as_documented
already covers the per-request key shape on the batch rows and the tool descriptions; nothing in
the suite read the single-shot rows, so a whole control -- here, the abstention gate the tool
already runs on every field -- could fall off the table while the code accepted it and the
description mentioned it. A caller who trusts the docs table hand-rolls their own confidence
check on values, never learning the tool already nulls an unsure field.

How it was verified

$ PYTHONDONTWRITEBYTECODE=1 python -B tests/test_mcp.py
...
   PASS docs-mcp/args_row_laya_status_names_every_param         missing=[] cell='none'
   PASS docs-mcp/args_row_laya_route_names_every_param          missing=[] cell='`state`, `questions`, optional `model`, `task`, `lang`, `lang_guess`'
   PASS docs-mcp/args_row_laya_predict_names_every_param        missing=[] cell='`state`, `questions`, optional `model` (`auto`, `english`, `multilingual`, or `typed-decisions`), `task`, `lang`, `lang_guess`, `max_len`, `head_max_len`, `min_confidence`'
   PASS docs-mcp/args_row_laya_shortlist_names_every_param      missing=[] cell='`state`, `questions`, optional `model`, `k` (default `20`), `task`, `lang`, `lang_guess`, `max_len`, `head_max_len`, `min_confidence`'
   PASS docs-mcp/args_row_laya_preset_names_every_param         missing=[] cell='`preset`, `state`, optional `task`, `lang`, `lang_guess`, `max_len`, `head_max_len`, `min_confidence`'
   PASS docs-mcp/args_row_laya_decide_names_every_param         missing=[] cell='`state`, `schema`, optional `model`, `min_confidence`'
   PASS docs-mcp/args_row_decide_drops_the_three_param_version
...
931 passed, 0 failed
all mcp tests passed

$ python -B laya-bench/mutate_cli_mcp_tool_args.py
baseline: OK (931 passed, 0 failed)

m1 restore the pre-fix three-param laya_decide row                     CAUGHT          FAIL docs-mcp/args_row_laya_decide_names_every_param  missing=['min_confidence']
m2 drop `min_confidence` from the laya_decide row (leave state/schema/model) CAUGHT   FAIL docs-mcp/args_row_laya_decide_names_every_param  missing=['min_confidence']
m3 drop `task` from the laya_route row                                 CAUGHT          FAIL docs-mcp/args_row_laya_route_names_every_param   missing=['task']
m4 drop `lang_guess` from the laya_shortlist row                       CAUGHT          FAIL docs-mcp/args_row_laya_shortlist_names_every_param missing=['lang_guess']
m5 drop `min_confidence` from the laya_preset row                      CAUGHT          FAIL docs-mcp/args_row_laya_preset_names_every_param  missing=['min_confidence']
m6 rename `min_confidence` param in laya_decide_tool (handler drift)   CAUGHT          FAIL docs-mcp/args_row_laya_decide_names_every_param  missing=['abstain_below']

6/6 caught

$ ruff check laya/ --select=E9,F63,F7,F82,F401,F811 --line-length=120
All checks passed!

$ python -B -m compileall -q laya/ tests/
(clean)

$ git diff HEAD~1 -- docs/cli-mcp.md tests/test_mcp.py \
    | grep -inE "(api[_-]?key|secret|token|password|bearer|sk-[a-zA-Z0-9]{10,}|ghp_|xox|BEGIN.*PRIVATE)"
(no hits)

Checklist

• Focused on one change (laya_decide row + its gate)
• Rebased on the latest main
• Tests pass locally (python -B tests/test_mcp.py -- 931/0)
• Docs or examples updated when the public API changed (docs page corrected; public API
  unchanged -- this PR makes the table describe what the handler already accepts)
• Mutation harness: laya-bench/mutate_cli_mcp_tool_args.py, 6/6 CAUGHT

[NandhaKishorM]

Merged in 0.3.29.

laya_decide accepts min_confidence and the docs row did not list it, so a caller trusting the table hand-rolls the abstention check on values and never learns the tool already nulls an unsure field. That is the specific cost of an incomplete parameter row, and it is worth more than the typo it looks like.

Your point about the "Main inputs" cell being a fourth copy of the shape, after the validator error message, each tool's registered description and the README, is the structural observation behind the whole batch. Four copies of one contract means three of them are wrong eventually. Reading the handler's own signature in the gate rather than asserting the corrected row is what makes this the last time.

Merge note: this conflicted with #984, which rewrote the laya_route_batch row in the same table to name its six keys. Both rows are kept, each PR's own, since taking one side whole would have reverted the other. Your test-side change landed as two separate additions, the gate and its entry in the runner list. tests/test_mcp.py is 957 of 957 on the merged result.