[codex] unify cache capacity config

[easel]

Summary

This is the org-visible stacked PR for the unified cache config model. It depends on #380.

Because the #380 head branch currently lives in the easel fork and this account cannot push branches to Luce-Org/lucebox-hub, this PR is temporarily opened against main. That makes it visible to the team, but the GitHub compare includes #380 plus the unified-cache commit. Once #380 lands, this PR should be rebased/retargeted to main so the visible diff collapses to the unified-cache work only.

What changed

• Replaces cache slot sizing with byte-sized RAM/disk budget flags for prefix and prefill caches.
• Keeps legacy slot flags as compatibility aliases.
• Adds disk-backed exact prefill cache support alongside existing prefix disk cache support.
• Exposes unified cache budget/usage telemetry in /props.
• Updates docs, OpenAPI props, entrypoint env vars, scripts, and cache proof tests.

Validation

• cmake --build server/build --target test_server_unit dflash_server -j$(nproc)
• server/build/test_server_unit (1959 assertions, 0 failures)
• python3 -m py_compile for changed Python scripts
• bash -n server/scripts/entrypoint.sh
• git diff --check
• DFLASH_SERVER_BIN=server/build/dflash_server python3 server/scripts/test_prefill_cache.py
  • RAM prefill cache active with 2 hits and ~10427x lower-bound warm speedup
• DFLASH_SERVER_BIN=server/build/dflash_server python3 server/scripts/test_prefill_disk_cache.py
  • disk prefill cache active with 2 hits and ~10623x lower-bound warm speedup

Notes

The cleaner branch layout would be base Luce-Org:codex/prefill-cache-wiring and head Luce-Org:codex/unified-cache-config, but pushing those branches requires org write permission.

[easel]

Claude Code reviewed the stacked diff from #380 head to this branch. It found stale /props contract docs and an OpenAPI prefix_cache example mismatch, plus a note that legacy full_cache.enabled is RAM-only in disk-only prefill mode. Addressed those in 5a321eb (docs: align props cache contract). No runtime correctness findings were reported in that review.

[easel]

Follow-up pushed in 7568ed2 to make the unified cache model the primary user surface. Defaults now use --cache-ram 1GiB split as 256MiB prefix + 768MiB exact prefill, and --cache-disk 16GiB split as 4GiB prefix + 12GiB exact prefill when a cache dir is configured. The server also alternates cold-miss snapshot targets when both RAM pools are viable so exact repeated prompts and multi-turn prefix reuse can both populate without user tuning.\n\nValidation after this follow-up:\n- cmake build: test_server_unit + dflash_server\n- server/build/test_server_unit: 1978 assertions, 0 failures\n- py_compile for touched Python scripts\n- bash -n server/scripts/entrypoint.sh\n- OpenAPI YAML parse/cache example assertion\n- git diff --check\n- RAM exact-prefill proof: 1 commit, 2 hits, warm prefill rounded to 0.000s\n- Disk exact-prefill proof: RAM off, 1 disk save, 2 disk hits\n\nClaude Code reviewed this follow-up diff and reported no actionable correctness bugs. It flagged only a cosmetic /props example indentation issue, fixed before this commit.