D — per-corpus background-work & defaults parametrization (registry v2 + ir maintain)

[thorwhalen]

Summary

There is no single place to parametrize a corpus's background work and its
build/storage/segmentation policy, nor smart defaults for new corpora. Today the
registry stores only {kind, embedder, params} and source_from_entry() cannot even
persist a strategy — it always reconstructs the preset default, so per-corpus
chunk_size, segmentation choices, storage backend, and any maintenance cadence are
unrepresentable. This issue adds a declarative per-corpus policy layer (registry v2) +
a defaults-per-kind rule set + an idempotent maintenance command.

Boundary respected: ir holds the declarative policy as data and exposes an idempotent
ir maintain; it does not run a scheduler. Execution (cron/launchd, or raglab's
budget governor / run-log per ADR #43) stays external. No Settings singleton — policy
is per-corpus data injected at build/maintain time (DI principle, semantic_search_design_notes §4.3).
ir must not import raglab (one-way dep).

Gap

• ir/registry.py: entry = {kind, embedder, params}; source_from_entry drops strategy.
• No per-corpus: segmentation/strategy spec, storage backend choice, reindex condition,
  synopsis policy, or "downtime" schedule.
• No documented defaults-per-kind (the three presets hardcode their own defaults).

Plan

1. Registry v2 schema (back-compatible). Extend the entry to optionally carry:

   {
     kind, embedder, params,
     strategy:   {name, params}        # persist & reconstruct the IndexingStrategy
     storage:    {backend, params}     # default "local" file store; seam for #28 / packed-matrix
     maintenance:{ reindex: {on: "source-change"|"interval", every: "24h"|null},
                   synopsis:{ enabled: false, scope: "recent"|..., window: "30d",
                              window_hours: ["02:00-06:00"] } }
   }
   

   v1 entries (no strategy/storage/maintenance) keep working unchanged.
   source_from_entry learns to reconstruct a persisted strategy.
2. Defaults-per-kind registry. A small pluggable table kind -> recommended {strategy, embedder, storage, maintenance} so registering a new corpus (or a new
   kind) starts from sensible, documented defaults instead of scattered constructor
   defaults. New kinds register their defaults here.
3. ir maintain [name|--all] — reads each corpus's maintenance policy and does the
   due work idempotently: incremental rebuild when the reindex condition is met, and
   (only if synopsis.enabled and within the configured window) refresh synopses on the
   in-scope slice. Safe to call from cron every N minutes; it no-ops when nothing is due.
   Ships with a documented cron/launchd snippet (the "downtime hours" live in the policy;
   the scheduler just calls ir maintain often).
4. CLI/register surface to set these without hand-editing JSON, plus ir info
   showing the resolved policy.

Why this shape

• Keeps the light path one line (ir build skills) — all of the above is optional with
  smart defaults.
• Makes background work declarative and inspectable (the "clear place" asked for),
  while keeping the runner external/idempotent so it composes with cron now and with
  raglab's observability/budget layer later (ADR ADR: one linked-artifact substrate, three operators (traverse / expand / purpose-index) #43) without ir importing raglab.

Related: #1 (roadmap), #38 (epic — raglab consumes ir's registry view), ADR #43
(scheduling/run-log boundary), #28 (storage backend seam), #C (first consumer:
sessions corpus segmentation + recent-only synopsis-in-downtime).

[thorwhalen]

Implemented

• ir/policy.py — MaintenancePolicy (reindex + synopsis), ReindexPolicy,
  SynopsisPolicy; DEFAULTS_BY_KIND smart rules (small enumerable kinds →
  source-change; sessions → interval/24h + synopsis-off + downtime [2,6]);
  resolve_policy (entry over kind over global), is_reindex_due, in_downtime
  (midnight-wrap aware), resolve_storage (only local; vd backend → Evaluate unifying ir storage behind vd.Collection (from ir_04) — DEFERRED #28).
• Registry v2 (back-compatible) — entries now persist optional strategy
  ({name,params}), maintenance, storage. Closes the concrete gap:
  source_from_entry now reconstructs a persisted strategy (it previously
  always reverted to the preset default). v1 entries resolve to their kind's
  smart default unchanged. Strategy (de)serialization in ir/strategy.py
  (STRATEGY_REGISTRY, strategy_to_spec/strategy_from_spec).
• ir/maintenance.py + ir maintain [name|--all] [--dry-run] — idempotent:
  incremental rebuild when reindex is due; when synopsis.enabled, the rebuild's
  strategy is with_synopsis-wrapped and run only inside the downtime window.
  Records last_maintained. Ships a cron snippet in the module docstring. ir runs
  the work; the scheduler stays external (no raglab import, no Settings singleton).
• ir info shows the resolved policy + last-maintained; ir register gains
  --reindex-on/--every-hours/--synopsis.

Tests: tests/test_policy.py (16). Full suite green. Constraints respected: DI (no
singleton), progressive disclosure (light path unchanged), one-way ir→raglab.