perf(core): bucket and index timeline projections

[hmans]

Why

Room Timeline retained decoded protobuf payloads and string-heavy metadata for
every room and every point in history. Threads independently duplicated many of
the same stable IDs in maps and summaries. Their RAM cost therefore grew with
the complete retained message history.

Closes #1721.

What changed

• Partition Room Timeline payloads into fixed per-room UTC-week buckets while
  preserving its one ordered EVT consumer and lightweight global indexes.
• Keep a configurable recent window decoded in RAM
  (core.room_timeline_hot_window /
  CHATTO_CORE_ROOM_TIMELINE_HOT_WINDOW, default 30d).
• Reconstruct cold buckets from exact EVT sequence references with 16 bounded
  direct reads, atomic installation, shared concurrent success/failure results,
  and context-aware errors.
• Select paginated thread references before hydration so opening one page of a
  long-lived thread does not load its complete history.
• Intern stable event, room, and user IDs once per projection, then represent
  timeline rows, posting lists, body state, echo/attachment indexes, thread
  mappings, summaries, participants, and follows with 32-bit handles, dense
  slices, and packed flags.
• Keep Room Timeline and Threads handle spaces independent, so replay,
  readiness, snapshots, and failure remain independently owned.
• Store only stable IDs and sequences at snapshot boundaries. Restore rebuilds
  process-local handles, with canonical stable-ID ordering and byte-identical
  round-trip coverage.
• Add operator diagnostics, retained-heap benchmarks, cold-load/race/snapshot
  tests, ADR-056, architecture/FDR updates, and operator configuration docs.

Results

• The deterministic 50k-message benchmark drops cold Room Timeline from 1,251
  to 276 retained heap bytes/message across the full PR (77.9%).
• Compared with the already-bucketed branch, compact indexing drops cold Room
  Timeline from 432.2 to 276.0 bytes/message (36.1%), Threads from 311.3 to
  118.1 bytes/message (62.1%), and their combined retained metadata from 743.5
  to 394.1 bytes/message (47.0%).
• A snapshot-disabled cold boot against the imported production-derived
  53,763-event history replayed Room Timeline in 1.14 seconds and Threads in
  1.14 seconds.
• On that same history, whole Go heap after forced GC fell from 45.14 MB on the
  already-bucketed branch to 34.66 MB with compact indexes (23.2%). The
  projection-specific diagnostics were also corrected to account for compact
  slice capacity instead of charging map overhead per element.
• Historical buckets remain resident after their first read in this version.
  Access timestamps and resident-byte accounting prepare a later bounded LRU.

Compatibility and rollout

• No public API, realtime protocol, or persisted EVT contract changes.
• The bucketed Room Timeline snapshot uses its schema-fingerprinted v2
  contract. The indexed representation preserves that stable protobuf shape;
  handles never cross the snapshot boundary.
• Threads preserves its existing v1 snapshot contract and likewise rebuilds
  its independent handles during restore.
• Pre-bucketing Room Timeline snapshots use a different contract fingerprint
  and safely cold-replay EVT once. Rolling back uses the old binary's
  independent contract namespace.
• Privacy boundaries are rechecked after lazy loads, so a concurrent
  retraction or user-key shred cannot reintroduce or return deleted content.

Test plan

• mise codegen-proto
• targeted Room Timeline, Thread, handle-table, and snapshot tests
• targeted race tests for cold hydration, snapshots, and adversarial
  handle-order restore
• repeated byte-identical snapshot round trips
• mise test-cli
• mise lint
• mise license-check
• architecture-inventory source-link and paragraph-length validation
• snapshot-disabled replay and forced-GC measurement against 53,763 retained
  production-derived EVT messages
• deterministic retained-heap benchmark at 50,000 logical messages / 124,478
  events
• adversarial code-review loop; both findings fixed, no remaining findings