Currency notice: count a practice skill behind the client and a screened practice reference that changed - #1136

…ned practice reference that changed

The command center ruled on B3 (2026-10-15): the daily currency run
triggers the lane's maintenance by events, not dates.

- practice_client_behind (W2): offline, from the fresh surface-watch
  report. It is 1 when versions.claude_code.watched is newer than the
  practice skill's client_version (release numbers compared as numbers)
  and starts the skill's step 2. A stale report or a missing skill gives
  no count; an unreadable version is "unknown".
- practice_refs_ahead (W2, under --network): practice-references.json
  takes an optional per-reference screen_paths. practice_references.py
  reads the changed file names from the compare call it already makes and
  reports screen_paths_changed (0 for an identical pin; None when not
  observed, or when the API's 300-file list holds no match).
  currency_due counts the references with a nonzero value. Pin drift
  alone does not count: measured live, 23 of 26 references drift, and the
  best-practice guide's 134 commits since its pin changed practice text in
  5 files. A failed check or an unknown change is the gap "practice
  reference check incomplete". Only shanraisshan/claude-code-best-practice
  carries screen_paths for now (README.md, best-practice/*).
- stack-currency.service: ExecStart passes --network (the CC's ruling,
  in the template so host and template stay identical), and
  TimeoutStartSec rises to 2400, covering every check's bound. The
  flag-echo test runs with the skill check stubbed.
- W4: the release-tracking and practice-references rows record the
  triggers, the function-check rule for headless roles, and the
  mise-shims PATH host note. The snapshot label follows.

Tests: test_currency_due, test_currency_due_notice,
test_model_currency_propose and test_practice_references: 187 OK. With
main's script, tool, template and catalog: 3 failures and 180 errors (the
harness's default practice-references stub needs the new check). The
live tool reports the guide at screen_paths_changed 5, matching the
compare read.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…t as gaps, never "nothing due"

The local review of ecfe6de (review_output.json 3002c574) found two P2s:

- tools/sota-convergence/practice_references.py: a fetch failure (timeout,
  rate limit) returned before the screen_paths step, so a screened
  reference had no screen_paths_changed field, currency_due's
  practice_refs_findings() dropped it and reported the coverage "checked"
  with a zero count, which could delete an existing notice unobserved.
  classify() now sets screen_paths_changed to None (unknown) for a screened
  reference it could not fetch, so the coverage is "unknown", a gap.
- scripts/currency_due.py: a practice_client "unknown" (an unreadable
  slots.json, or a surface report with no valid watched version) was not a
  gap, so with every count at zero the run printed "nothing due" and
  deleted the previous due-file. It is now the gap "practice skill check
  incomplete" in both the run's gap list and gaps_of(), so incomplete()
  holds.

Fail-first: test_a_failed_fetch_of_a_screened_reference_is_unknown_never_dropped
and test_an_unknown_practice_client_is_a_gap_never_nothing_due failed on
ecfe6de ('screen_paths_changed' not found; gap list []); the practice and
currency modules run 166 tests OK; validate rc 0.

Co-Authored-By: Claude Opus 5.5 <[email protected]>