Workers: cover logging gaps so each one is visible at default log level

[sysread]

=head1 SYNOPSIS

Cover the three logging gaps an audit surfaced. Each background
worker now lands at least one .info line per cycle outcome that
matters, so all eight workers are visible at the drawer's default
log level.

=head1 PURPOSE

After shipping the wiki librarian, the user observed they
"weren't seeing the librarian fire" in the Logs drawer and asked
whether all the workers had consistent logging. They didn't.
Three real gaps:

1. attachment-expiry had no createLogger at all. The only
   log channel was the legacy post({type:'log'}) inside the
   worker entry, which only fires on setup-session failure and
   lease loss. So an attachment-expiry problem in production
   would leave no breadcrumb in the drawer.
2. samskara logged at .debug what the other agent workers
   log at .info. With the drawer at the default level, you saw
   every other worker fire but not this one.
3. wiki-librarian had logs on too-small / reviewed /
   error but not on acquired-lease / polling / too-soon.
   Since too-soon is the most common state for the first 12
   hours after a deploy (the cross-device claim won't return
   true until the configured interval has elapsed), the user
   couldn't tell whether the librarian was firing at all - it
   was silent rather than off.

=head1 DESCRIPTION

=head2 Layer 1: how logging looked across the 8 workers

Source-tag convention is consistent (every manager declares
<feature>-worker as loggerSource, except embed-worker which
keeps its abbreviated form intentionally). Five of the eight
workers (embed, reflection, journal, wiki, summary)
follow the same shape: .info on lifecycle headlines (picked up X /
finished X / errored on X), .debug on per-row details. The other
three diverge:

• attachment-expiry-worker: no createLogger. Loop is silent.
• samskara-worker: lifecycle events on .debug (was unique
  among the eight in this).
• wiki-librarian-worker: lifecycle covered for some outcomes
  (the work-happened path) but silent on the more common
  cooldown path.

=head2 Layer 2: what this PR changes, parallel to layer 1

attachment-expiry-worker gets a createLogger and four
lifecycle lines:

• .info on lease acquired ("lease acquired - starting expiry
  sweep").
• .info on each non-zero expire batch ("expired N attachment
  row(s) older than D days") - now you can see the worker
  actually doing work.
• .info on RPC failures.
• .debug on polling and empty-queue ticks so an idle worker
  isn't noisy.

samskara-worker has four lifecycle lines bumped from .debug
to .info:

• assimilate: claimed substrate <id> (thread <thread>)
• assimilate: saved substrate <id>
• pair-relate: selected pair <a> <> <b> (cosine X.XXX)
• pair-relate: associated <a> <> <b> (<kind>: <label>)
• compound-regen: synthesizing summary from N sample row(s) (cap M)

Detail / error / null-result lines stay on .debug so the drawer
isn't flooded with samskara's higher per-cycle volume.

wiki-librarian-worker gets logs on every cycle outcome:

• .info on lease acquired.
• .debug on polling (60s loop, would otherwise fire every
  minute).
• .info on too-soon, with the configured minimum interval
  named explicitly: "not yet eligible for a librarian run (min
  interval 12h since last successful run)". ~24 lines/day during
  cooldown gives the user the "yes, the librarian is alive and
  waiting" signal that was missing.
• .info on claim acquired (the transition from waiting to
  actually running).
• Existing .info lines on too-small / reviewed retained.
• Error paths that were silently .debug (claim RPC, list,
  agent threw) are bumped to .info so a real failure isn't
  silent.

Embed-worker is left as-is (the abbreviated embed- source tag
is intentional per the user's preference; existing log shape is
fine).

=head2 Layer 3: how that resolves PURPOSE

After this PR, opening the Logs drawer at the default level shows
every background worker firing on its own cadence:

• attachment-expiry on each expire batch.
• samskara on each phase pickup and write.
• wiki-librarian whenever it's awake, including the explicit
  "waiting on min-interval" line during the 12h cooldown.

So the user's "I'm not seeing the librarian fire" question is now
answered structurally: if the librarian truly is firing, you'll
see the .info too-soon or reviewed line. If you don't see
either, the worker isn't actually running - the lease isn't being
acquired, or the worker is disabled in Settings, or it crashed
without reporting.

=head1 Notes for AI reviewers

• The librarian's too-soon line lands every idle interval (1h)
  during cooldown. ~24 lines/day. This is intentional - the user
  asked for visibility specifically because they couldn't tell
  whether the worker was alive. Throttling it to "first too-soon
  after acquisition only" would be silent again on an unlocked
  laptop that runs continuously through a cooldown window.
• Samskara's bumps are bounded to lifecycle headlines (claimed
  / selected / associated / saved). Per-row details and per-cycle
  null-returns stay on .debug. The volume is bounded by the per-
  hour samskara claim cadence, not per-message.
• attachment-expiry's empty-queue line is .debug deliberately.
  Once the backlog drains, that branch fires hourly forever, and
  a heartbeat at .info would dominate the drawer for the long
  steady state.

https://claude.ai/code/session_015XcR7xzLdij66ZbYERUdLH

---

Generated by Claude Code