Proposal: entries are written once; keys are managed separately

[PaulSnow]

Design proposal (draft PR for review; stacked on #95 whose platform run is its evidence). On acceptance it becomes spec text in 1.5, 1.8, 2.3 and 2.7, and #94 leaves the register.

docs/proposals/2026-09-16-entries-written-once.md:

• The rule: an entry is written once; reclaiming space and bounding the walk are done on keys, never by rewriting entries. Entries move only to make a bigger hole, one per bounded step.
• Dynamic layer: a heap with holes. A rewrite that fits reuses its slot; one that doesn't appends and frees a hole; holes by size class; a hole is reusable one seal late; entries carry length and checksum so a torn slot is a hole, not a wrong answer.
• Permanent layer: append-only data file per shard; each block seals an index delta; merge and pack fold indexes (~40 B a record), never entries.
• Seal: one commit point per store per block (parallel data sync, one block file of index deltas, the manifest) instead of four barriers per shard. Closes A seal costs six fsyncs and rewrites the whole manifest; imports pay the same per segment #33.
• Acceptance: cmd/bdbench -stores 9 flat for 30 minutes: seal p90 under 100 ms, every block inside the interval, maintenance bytes near ingest, zero wrong answers. Today that run reaches seal p90 731 ms and misses 46 of 540 blocks by minute 3.

🤖 Generated with Claude Code

https://claude.ai/code/session_015wUZmWRdAPtcfFgYrLJCc5