Anvil trimming optimization for very large servers

[mattiapearl]

The idea is that long-lived region files can retain physically unused trailing sectors with no purpose. We can return this trailing space to the filesystem. It's not a form of general compression. NICHE optimization for extremely large servers.

Benchmarks

• Executor: Pi AI agent
• Direction and code review: contributor

Final Paper implementation

Actual RegionFile, 256 deterministic 32 KiB records, repeated grow/shrink operations across three paired runs:

Metric	Stock	Candidate
Full file	9,441,301 B	9,441,301 B
Trailing half cleared	9,441,301 B	4,726,784 B
Everything cleared	9,441,301 B	8,192 B
Interior deletion	9,441,301 B	9,441,301 B
Shrinking-flush median	1.4402 ms	1.3872 ms
Shrinking-flush p95	24.4105 ms	15.0366 ms
Clear-all flush median	0.4557 ms	1.2518 ms

Earlier five-pair Folia campaign

The same physical-space behavior was measured through actual RegionFile operations:

• Constructed trailing-half deletion reclaimed 49.94%.
• Clearing all records reclaimed 99.91%.
• Interior holes remained.
• Retained payloads survived flush, close, and reopen.
• The earlier implementation's shrinking-flush median was 9.10 ms stock versus 15.97 ms patched; the final implementation added a dirty-tail flag to avoid unnecessary truncation checks.

Fresh-world scan

• 1,128 .mca files
• 151,533 records
• 683,713,425 total bytes
• 0 files with reclaimable tails
• 0 reclaimable bytes

Final Folia integration smoke

• Full file: 9,441,301 B
• Trailing half cleared: 4,726,784 B
• Everything cleared: 8,192 B
• Interior-only deletion: 9,441,301 B
• Retained payload verification passed.

AI assistance disclosure

• AI materially assisted—not merely autocomplete.
• Primary model: gpt-5.6-sol, maximum reasoning.
• Seven gpt-5.6-terra:medium research agents.
• Two gpt-5.6-luna:low read-only scouts.
• Assistance covered research, code inspection, implementation, tests, benchmark harnesses, analysis, and issue/PR searches.
• No independent AI reviewer was used.
• The passive review ledger only recorded prompts/responses; it did not perform review.
• I reviewed all the code, understand it, and accept responsibility for the contribution.

[electronicboy]

Not fond of this being a core file patch, and generally concerned that this is just something that won't be able to be tested enough for us to guarantee its safety across updates, etc