feat(synthesize): file good answers back as page proposals

[philluiz2323]

What changed

kb.synthesize gains an opt-in file_as_page: bool = False parameter
(plus page_title: str | None = None). When set, the synthesized answer
is filed back as a PAGE proposal via the existing propose_page, citing
the same claim ids the answer already cites. Registered on all three
transports plus CLI: MCP kb_synthesize(..., file_as_page, page_title),
JSONL kb.synthesize params, vouch synthesize --file-as-page [--page-title T] [--agent A].

Why

ROADMAP.md's 1.4 milestone: "Good synthesized answers can be filed back
as page proposals — gated by review like every other write." kb.synthesize
already produces exactly that shape — citation-bearing prose traceable to
approved claims — but nothing wired it to propose_page; turning a good
answer into a page meant manually copying the text and re-typing the
cited claim ids into a separate call.

Closes #736

Still fully gated: propose_page files a pending proposal, the same as
every other write path in this codebase — nothing here approves anything.
Filing is skipped (not an error) when the answer is empty or cites
nothing — an uncited "answer" is the KB saying it doesn't know, not
knowledge worth filing — or when propose_page itself fails (e.g. a
title collision): the synthesis already succeeded, and losing that to a
secondary, optional step would be a worse failure than reporting why
filing didn't happen. Both outcomes surface via new page_proposal_id
(None when skipped or failed) and page_proposal_skipped_reason result
fields.

Internally, the existing no-LLM path is extracted into
_deterministic_synthesize so a new _maybe_file_page_proposal helper
can wrap either backend's result the same way — the LLM path already
produces the identical {query, answer, claims, pages, gaps, _meta}
shape, so one wrapper covers both without duplicating the filing logic.

What might break

Nothing — purely additive and opt-in. file_as_page defaults to False,
so every existing caller (MCP, JSONL, CLI) gets back the exact same
result shape as before; the new page_proposal_id/
page_proposal_skipped_reason keys are only present when file_as_page
was requested. No on-disk shape, object model, or kb.* method removal —
propose_page and the review gate itself are completely untouched.

VEP

Not applicable — no object model, on-disk layout, bundle format, or
audit-log shape change. A new opt-in parameter on an existing read method
that calls an existing, unmodified write path.

Tests

• Local make check-equivalent: ruff clean (src + tests); mypy
  clean; tests/test_capabilities.py, tests/test_cli.py,
  tests/test_hot_memory.py, tests/test_synthesize.py all pass
  (the 3 pre-existing LLM-backend test failures are a known Windows
  python3-stub-invocation environment issue, confirmed present
  identically on test HEAD via git stash comparison, unrelated to
  this change)
• New behaviour has tests — 8 new cases in tests/test_synthesize.py:
  files a pending proposal with the right claims/body/actor, custom
  title, skip-on-empty-answer, ValueError without proposed_by,
  default-off backward compatibility, plus JSONL and CLI
  (CliRunner) end-to-end cases
• CHANGELOG.md updated under ## [Unreleased]