core: partition _hooklib.py god module (architecture-002)

[SUaDtL]

Branch: feat/pi-support · Tribunal run 2026-07-17-pi-support-branch · Finding: findings/architecture/architecture-002.json
1,201 LOC / 46 functions / 28 of 44 consumers. Partition along concern seams (_activationlib, _sensitivelib). Sequence after architecture-001.

[SUaDtL]

Sprint backlog-burn-2026-07-17: deliberately deferred — partitioning _hooklib (28 consumers, 3 hosts) in the same window as the #320 pre-bash extraction doubles churn on the synced enforcement surface. Sequenced after #320 lands and bakes.

[SUaDtL]

Re-measured 2026-07-26 — still real, and growing

The finding was filed at 1,201 LOC / 46 functions / 28 of 44 consumers. Measured today on main:

	filed	now
LOC	1,201	1,221
functions	46	46
consumers	28 of 44	31 of 48

So it has not stabilised — three more modules took a dependency on it since the tribunal ran. Worth stating plainly because an earlier note in this campaign leaned toward closing this on the grounds that its acceptance criterion was unreachable. That criterion is not in this issue, and the measured trend does not support closing it.

The proposed seams are real, but they are not the big one

Grouping the 46 definitions by concern:

• path classification — norm_path, repo_rel, classify_protected, is_audit_log, is_tail_append, is_decisions_path, is_context_md, is_marker_path, is_migration_path, is_ci_path, is_deploy_path, path_in_globs, scope_globs, migration_globs, _glob_to_re, _scope_res — 16
• hook protocol and reporting — read_input, tool_input, block, remind, warn, _log_gate_event, staleness_warning, marker_fresh — 8
• activation — arbiter_active, frontmatter_enabled, frontmatter_enabled_text, project_root, _root_cache_key, _reset_root_cache — 6
• IO, atomics, locking — write_text_atomic, acquire_lock, release_lock, _is_lock_contention, utf8_stdio — 5
• sensitive scanning — is_sensitive_scan_exempt, diff_added_lines, sensitive_scan_added_lines, line_digest, content_digest — 5
• host seam (ADR-0011) — get_host, set_host, reset_host — 3
• controls config — _read_controls, _controls_mtime, _custom_re_cache — 3

The suggested _activationlib + _sensitivelib split moves 11 of 46. The dominant cluster is path classification at 16, and it is also the most self-contained — it is nearly pure functions over a string plus the glob cache, with no dependency on the host seam, the gate-event sink, or the lock.

What makes this expensive, and it is not the LOC

core/pysrc/ is CANONICAL and vendored to three plugin trees by tools/sync-core.py, so every new module lands in four places and each import rewrite is a payload change to ca, ca-codex, and ca-pi simultaneously — three version bumps and three changelog entries per slice. That is the real cost driver, and it argues for few, large slices rather than one-module-at-a-time: the same migration done in five passes pays the three-plugin tax five times.

Suggested shape, if this is taken

1. _pathlib-style module for the 16 path classifiers — biggest win, cleanest seam, no host coupling.
2. _activationlib + _sensitivelib as the tribunal proposed, folded into the same pass to pay the vendoring tax once.

That leaves _hooklib as protocol, reporting, IO, locking, and the host seam — roughly 25 functions, which is a plausible resting point rather than another god module in waiting.

Not started; this needs an explicit call on appetite, since it is a wide mechanical refactor across 31 consumers × 4 trees with no user-visible behaviour change.