v1.79.1.0 Make setup, sessions, and GBrain sync fail closed - #2758

Two hooks, both pull-mode, because gbrain's push lane needs a single-process
PGLite engine and a shared brain cannot be one.

session-start-hook puts the governing state in front of the agent before it
guesses: the active goal, work that only exists on this disk, and the pointer
to the factory map. On 2026-09-01 and twice on 2026-09-03 agents opened a
session, read a stale file instead of the running system, and handed the CEO
confident wrong diagnoses. The brain held the right answer every time and
nothing made them look.

skill-router-hook matches the user's own words against every skill's declared
triggers on each message and names the skills that own the request. The
routing contract already says to do this by hand; an agent that forgets to
look routes on vibes. Triggers are declared in English, so the matcher also
carries the Spanish phrasings this user actually types — "ponte al día",
"se rompe", "vale la pena" — widening the haystack, never replacing his text.

Both stay silent when nothing matches, never block a turn, and never spend on
a model. They surface candidates; the agent still decides.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

docs/operations holds twenty-three memos and the open issues carry the
decided design. On 2026-09-04 a session spent five hours rebuilding, from
arxiv, a conclusion the 2026-07-08 coverage-ontology memo had already reached
from the same sources: SBA, HBS/Sahlman, Business Model Canvas, NAICS, Steve
Blank, Lean Startup, W3C PROV, FAIR, GOV.UK. The work was not missing. Nobody
looked, because nothing said it was there.

The start hook already named the factory map. It now also names the memo
directory with its file count and tells the agent to search open issues before
concluding a design is undecided.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…ain-readiness