refactor(core): extract activation from _hooklib, completing the partition (#321, slice 4/4)

[SUaDtL]

Last of four. _hooklib finishes at 557 lines, from 1,263 — a 56% reduction — and #321's architecture-002 finding is closed.

module	lines
_hooklib	557
_sensitivelib	270
_scopelib	262
_protectedlib	204
_activationlib	183
_pathnorm	44

What moved

_activationlib owns the three answers every hook needs before it can decide anything: which Host this process runs under, where the project root is, and whether the arbiter is switched on. They share two process-lifetime caches — splitting them would put a cache in one module and its only writer in another.

The edge runs the other way for the first time: _log_gate_event stays in _hooklib and calls get_host()/project_root(), so _hooklib imports from _activationlib and never the reverse. Still one-way, still no cycle.

The mutable global was the real hazard

It is why this slice went last.

_HOST is deliberately not re-exported. from _activationlib import _HOST binds the value at import time, so a later set_host() would rebind the name in the owning module and leave _hooklib holding a stale None forever — a silent divergence in the host seam, which is the worst possible place for one.

The three accessors are re-exported instead, because a function binding stays live. Proven rather than reasoned:

set via _hooklib       -> visible in _activationlib: Fake
set via _activationlib -> visible in _hooklib:       Fake
reset via facade cleared the owner global: True
SHARED-CACHE CONTRACT HOLDS

That is the exact opposite of slice 3's lesson, and the distinction is precise: re-export a private function, never a private mutable value.

A methodology bug I should flag

Three times this campaign I concluded a symbol had "no external consumers" and was wrong on contact — reset_host, seven scope privates, and _reset_root_cache.

One cause: my exclusion pathspec :!*_hooklib.py also matches test_hooklib.py — the very file that consumes these symbols through _hooklib.. So the check silently excluded its own best evidence.

Corrected here by excluding the module paths explicitly rather than by suffix glob. On that corrected evidence _reset_root_cache is re-exported; _HOST, _ROOT_CACHE and _root_cache_key are genuinely unreferenced.

Verification

Public surface against origin/main: nothing gained, and the only losses are those three verified-unreferenced privates.

hook test tree	1185 pass
test_hook_guards	176 assertions, 0 failed
test_hooks_cold_install	285 assertions, ca + ca-codex, REAL/STUB/PY2/NONE
test_codex_adapter, test_dual_host_store	pass
sync-core --check	53 core files × 3 plugins byte-identical

Parity rests on tests that did not move — none were edited in any of the four slices.

Versions

ca 2.10.7 → 2.10.8, ca-codex 0.3.6 → 0.3.7, ca-pi 0.1.42 → 0.1.43.

Closes #321.