v0.0.4 - Crash harness & failpoint hooks

What's new

BeachDB now has a structured crash testing harness and an internal failpoint framework for exercising durability under process crashes and injected engine faults.

Crash harness (cmd/crash)

Replaces the previous kill-loop orchestrator with a controller/worker architecture:

• Deterministic workloads — pre-generated from a seed, fully reproducible
• NDJSON protocol — binary-safe controller↔worker communication over stdin/stdout
• Oracle verification — after every crash/reopen cycle, recovered state is checked against all acknowledged operations
• Replayable artifacts — each run writes a JSON artifact containing workload, events, cycle metadata, and any verification failure
• Subcommands: crash run, crash replay, crash worker

Failpoint framework (internal/crashhook)

Environment-variable-armed crash and fault injection hooks wired into the engine:

• Crash points: wal_after_append, wal_after_sync, flush_after_file_sync, flush_after_publish
• Fault points: wal_sync_error, sst_write_error, sst_publish_error

All hook sites are marked with // FAILPOINT: <name> for grep-ability.

Other changes

• make crash-check runs a short deterministic smoke run
• docs/testing.md updated with harness and fault injection documentation
• Engine-level subprocess tests for crash hooks (engine/crashhook_test.go)