Weather questions currently traverse the two-level orchestrator/worker LLM
agent loop, taking 21-63s. This adds a deterministic fast-path that intercepts
weather intents BEFORE any agent/LLM runs, answers directly from the Open-Meteo
HTTP API, and returns in ~300-500ms. Anything that does not cleanly match
falls through untouched to the normal agent pipeline.

New module intent_fast_path.py (repo root, added to pyproject py-modules):
  - register_intent()/_intent_fast_path() registry so future intents are one
    line. Any matcher/handler exception is swallowed -> fall through.
  - End-anchored weather matcher that ignores conversational prose
    ("weather affects my mood") and slash commands.
  - httpx.AsyncClient(timeout=2.0) handler. Default location is hardcoded
    Woodstock, IL (no geocoding). Named locations are cleaned of ", ST ZIP"
    before geocoding (Open-Meteo geocoder chokes on those). WMO codes mapped
    to short text; terse Fahrenheit current + 3-day forecast, Telegram-safe.
  - STRICT FALL-THROUGH: returns None on no match, empty geocoding, timeout,
    HTTP 4xx/5xx, JSON/parse error, or missing current.temperature_2m. Never
    returns an empty/partial/wrong string.

Two insertion points, both behind a guarded import (missing module = safe
no-op stub that always defers to the agent):
  - gateway/platforms/api_server.py _handle_chat_completions: after message
    validation, honoring the stream flag (OpenAI chat.completion JSON or SSE
    chunks + [DONE]); CORS via _cors_headers_for_origin; INFO log on hit.
  - gateway/run.py GatewayRunner._handle_message: between the Telegram lobby
    block and the session-sentinel claim, so the session lock is NOT taken for
    a fast-path reply. Downstream of auth + mention-gating, which are preserved.

Tests: tests/test_intent_fast_path.py (34 cases) cover matcher positives/
negatives, location cleaning, handler fall-through (timeout/500/empty
geocoding/missing temp), success render, and dispatch exception safety. No
network in the unit suite (httpx monkeypatched). 34 passed; ruff clean.

Tradeoff: fast-path replies are not written to the session transcript
(history bypass) — acceptable for ephemeral weather answers.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… review

Adversarial review of the weather fast-path found the connector path
("in|for|at|near|around <X>") accepting ANY trailing noun as a location,
so non-weather prose got geocoded into a bogus weather answer (CONFIRMED
LIVE: "forecast for the meeting" -> Nenagh, Ireland; "forecast in the lab"
-> Indiana). It also let a noun-filler bridge to a connector
("weather report for the Q3 sales").

HIGH-1: Add _connector_location_ok() guarding loc1/loc2 connector captures
  in both _weather_matcher and _extract_location — reject "the <noun>",
  >5 tokens, >60 chars, or any _NON_PLACE_WORDS stopword (meeting, lab,
  code, budget, sales, quarter, report, market, stock, project, team, …).
  Tolerates ZIP/state tokens so "Woodstock, IL 60098" still passes.
HIGH-2: Restructure _WEATHER_RE into two end-anchored, non-optional
  alternations: (a) keyword + optional filler + EOL, or (b) keyword +
  optional idiomatic "like" + connector + location + EOL. A noun filler
  (report/budget/…) can no longer be followed by connector+location.
  ("what's the weather like in Denver" preserved via the "like" bridge.)
MEDIUM: Replace flat httpx timeout=2.0 (per-call, ~4s worst case) with
  httpx.Timeout(connect=1.0, read=1.5) and wrap the two-call named-location
  branch in asyncio.wait_for(..., 3.5s) hard ceiling. Default Woodstock
  path stays single-call.
LOW: _is_place_like rejects single tokens >60 chars.

Tests: +20 cases (54 total) — all confirmed-live false positives asserted
no-match AND _extract_location None; true positives preserved; named-branch
hard-ceiling test with a mocked slow httpx. ruff clean; full fast-path suite
and 150 gateway api_server tests green.

SSE (api_server): split the fast-path role+content single delta into
spec-standard chunks (role-only, content, finish_reason:stop, [DONE]) so
strict OpenAI/OpenWebUI stream consumers parse cleanly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>