fix: land grandcamel's four wire-shape fixes from an in-repo branch

[anthmFS]

Lands #84, #85, #86 and #87 from an in-repo branch so they can actually run CI, plus the one line worth keeping from #83.

Why re-homed rather than merged directly: those four are fork PRs, and fork
PRs do not receive repository secrets. Every job died in ~6s on
Input required and not supplied: token at the Checkout porting-sdk step,
which needs PORTING_SDK_TOKEN. They can never go green from a fork, whatever
is in them. Commits are cherry-picked with -x, so grandcamel stays the author
of all four.

The four, each verified against the authoritative source

#87 — execute_swml transfer rides beside the SWML document. Confirmed in
the engine: mod_openai/actions.c:129-140 reads transfer off item (the
action object), not off thing (the document). Inside the document it is never
read, so do_transfer stays 0 and execute_swml(..., transfer=True) does
nothing today. connect() and swml_transfer() already emit the sibling shape;
this was the outlier. Note actions.c also gates transfer on from_relay, so
the behaviour change only shows on relay-originated calls.

#86 — tap direction is speak/listen/both. The engine-generated SWML types say
direction: Literal["speak"] | Literal["listen"] | Literal["both"] and
codec: Literal["PCMU"] | Literal["PCMA"]. The SDK's runtime validator accepted
"hear" (which the engine rejects) and refused "listen" (the correct value).
It also omitted direction when it equalled the SDK default "both" — and the
SWML verb defaults to "speak" when absent, so tap(uri) asked for both and got
speak-only. Now always emitted.

#84 — conversation_timeout forwarded on the start path. prepare() puts it
in the params (gateway.py:451) and create_conversation() accepts timeout=,
but the HTTP route rebuilt the call and dropped it, so the browser was told 900
while the service kept its 3600 default.

#85 — live_transcribe docs + ai_sidecar test. The generated REST client takes
action: LiveTranscribeStartAction | … (object keyed by phase); the docs showed
action="start", lang="en". And ai_sidecar is in the bundled schema now, so
the test's raw-document bypass is obsolete — it goes through
add_verb_to_section and gets schema validation.

My two follow-up commits

The conversation_timeout test arrived unannotated, which reds TYPECHECK (mypy
has tests in scope on purpose). Annotated. And from __future__ import annotations in search_service.py is the one line salvaged from #83, which was
otherwise superseded by #81 — measured on its own it changes no mypy count, so
it is hygiene.

Verification

Full run-ci: everything passes except TYPECHECK and TEST, both at main's exact
baseline — 19 mypy findings in mcp_gateway/search that depend on which optional
extras are installed and that CI does not report, and the 6 pre-existing
mcp_gateway test failures that fail identically on unmodified main. 5942 tests
pass (up 2, the new ones).

Heads-up: #86 and #87 change emitted wire shapes

Signatures barely move, so DRIFT will not notice, but EMISSION byte-compares
to_dict() and the 9 ports still emit the old shapes. This adds port work on top
of the ChatGateway drift. Both are fixing silently-broken behaviour, so the risk
is code that has been relying on the breakage — chiefly tap() now sending both
directions instead of speak-only.

Closes #84. Closes #85. Closes #86. Closes #87.

🤖 Generated with Claude Code

https://claude.ai/code/session_015dYktt85Ltj3oK9gG5VBww