Simplify file cache metadata and CI benchmarks

[joecorall]

Summary

This PR simplifies the file-backed cache paths and trims the benchmark harness to measure Triplet only.

Main changes:

• move cache size/age accounting to payload file stat/mtime
• keep source-cache content type sidecars, but remove derivative-cache content-type sidecars
• stop mutating derivative cache entries on read hits
• harden cache eviction against concurrent writes and temp-file races
• make benchmark-iiif Triplet-only instead of spinning up Cantaloupe

Why

Cached benchmark latency was still much higher than expected. The first cleanup was reducing cache-hit filesystem overhead:

• no read-side mtime refresh
• no derivative .meta reads on hit
• derivative cache serves content type from request format

The benchmark harness also made local/CI analysis noisier by always benchmarking Cantaloupe in the same run. This PR removes that and makes the benchmark output Triplet-focused.

Changes

Cache behavior

• FileStore now derives:
  • Entry.Size from payload stat
  • Entry.StoredAt from payload mtime
• max_age and max_bytes eviction now use payload file mtime
• derivative cache uses a payload-only file store
• source cache still stores content type sidecars

Derivative cache hits

• no sidecar read on normal derivative cache hit
• no cache writes on normal derivative cache hit
• handler falls back to IIIF request format for Content-Type

Concurrency hardening

• temp files are skipped during eviction
• cache mode is constructor-only, not mutable after creation
• added tests for:
  • payload-only derivative entries
  • mod-time eviction behavior
  • fake-time expiry behavior with synctest
  • in-flight temp-file preservation
  • concurrent Put() calls under constant eviction pressure

Benchmarks

• make benchmark-iiif now runs Triplet only
• benchmark report metadata no longer assumes Cantaloupe is present

Notes

• If an environment still has old derivative .meta files from before the payload-only derivative store change, clearing the derivative cache is the cleanest reset.

speedup

Ran the benchmarks locally on the old codebase and this PR

make benchmark-iiif

before

Mode	Concurrency	Triplet OK	Duration s	Req/s	p95 ms	p99 ms	CPU ms/req	Max MiB
uncached	2	270/270 (100%)	18.16	14.9	156.9	287.0	53.77	138.1
uncached	4	270/270 (100%)	10.20	26.5	193.7	400.6	43.08	229.0
uncached	8	270/270 (100%)	7.77	34.7	309.0	474.8	62.82	298.9
cached	8	270/270 (100%)	4.06	66.5	11.1	22.8	1.41	129.5
cached	32	270/270 (100%)	3.98	67.9	194.5	256.0	3.72	124.9
cached	128	270/270 (100%)	3.76	71.7	323.5	444.4	2.99	142.7

after

Mode	Concurrency	Triplet OK	Duration s	Req/s	p95 ms	p99 ms	CPU ms/req	Max MiB
uncached	2	270/270 (100%)	18.17	14.9	153.4	278.8	57.89	170.4
uncached	4	270/270 (100%)	10.01	27.0	184.6	396.8	41.14	228.2
uncached	8	270/270 (100%)	7.36	36.7	295.2	477.6	57.57	320.0
cached	8	270/270 (100%)	3.75	72.0	10.3	23.3	1.29	140.4
cached	32	270/270 (100%)	3.59	75.2	188.8	240.4	3.96	153.3
cached	128	270/270 (100%)	3.63	74.5	320.7	372.2	2.84	150.2

[github-actions]

Benchmark Matrix: pr-33-25251913018

Summary

Triplet image: ghcr.io/libops/triplet:cache-eviction-storedat

Mode	Concurrency	Triplet OK	Duration s	Req/s	p95 ms	p99 ms	CPU ms/req	Max MiB
uncached	2	270/270 (100%)	13.32	20.3	170.6	297.9	80.09	372.0
uncached	4	270/270 (100%)	10.16	26.6	271.4	472.3	72.99	172.3
uncached	8	270/270 (100%)	10.10	26.7	404.7	621.4	73.26	270.3
cached	8	270/270 (100%)	10.18	26.5	408.5	674.8	74.90	247.7
cached	32	270/270 (100%)	10.55	25.6	558.4	715.8	74.28	341.8
cached	128	270/270 (100%)	10.71	25.2	886.0	1049.1	78.31	410.7

Status reflects Triplet request success. Performance metrics are informational.