bedrock-kv / bedrock / 734faabad0557b8c84ffd93a3a96dd7d7607b876

82%

Build Type

push

github

Committed by web-flow

Commit Message

Roll the WAL on the Demux's configured cut interval, not the default (#212)

Closes bedrock-61c.1 (epic bedrock-61c).

The WAL's segment-roll boundary and the Demux's cut boundary are the
**same boundary** — a segment holds exactly one cut bucket, which is
what lets trimming drop history at the cut cadence even though the
active segment is trim-immune. Both modules document this.

They agreed only by coincidence:
- `Pushing.crosses_cut_boundary?/2` read
`Demux.Server.default_cut_interval_us/0` directly
- `DemuxControl.start/1` — documented as the single place a log's Demux
is started from, where *"initialization and recovery reset must agree on
every option"* — never passed `:cut_interval_us` at all

Both sides defaulted, so both matched, and the Demux's documented
`:cut_interval_us` option was unreachable from a log.

### Changes
- `State.cut_interval_us/1` is the single resolution point for the
default
- `DemuxControl` hands that value to the Demux; `Pushing` measures the
roll against the same value
- carried on the **manifest params**, not `State` alone: the Foreman
rebuilds a restarted worker from its manifest, so an init-only field
would revert to the default after a crash and the log would resume
rolling on boundaries its already-written chunks were not cut on —
reintroducing across a restart exactly the drift this closes in-process

### Scope
This is a coupling fix. No production behaviour changes today, because
until now nothing could configure the width. A drift's cost is
**retention, not durability**: a segment spanning several cuts stays
un-trimmable until the last cut covering it lands. The trim floor is the
Demux watermark alone and eligibility is content-derived, so an
unaligned segment is never trimmed early.

### Verification
2802 tests, 0 failures. Credo and dialyzer clean.

Adversarially audited by a cold agent, which found no off-by-one at the
bucket boundary, confirmed no explicit `nil` can reach `div/2`, verified
the recovery-rese... (continued)

Coverage Stats

7 of 8 new or added lines in 3 files covered. (87.5%)

12 existing lines in 1 file now uncovered.

6796 of 8652 relevant lines covered (78.55%)

1027.78 hits per line