[niemyjski]

Stacked on #582. Preserve live cache entries when incoming values expire during preparation, reclaim expired stored entries through the atomic update path, and reduce unnecessary list-update allocation while preserving existing custom-calculator contracts.

Validation: Current head 143f24fe passed both hosted build/test runs. Local core validation at ca9b5c9a passed Release build/formatting and 2,446 tests with 22 skips; the only subsequent source change is the sample-only advisory correction, with src/ and tests/ unchanged. Push CI and PR CI passed. The expired-entry review finding is resolved.

Combined validation: 2,484 core tests and 778 Redis cache/lock tests passed; 22 core and 18 Redis existing skips; zero failures. Binary API checks passed for core net8/net10 and TestHarness net10. All-provider source build passed with an isolated RabbitMQ AppHost dependency pin; unmodified RabbitMQ main still fails its MessagePack advisory check.

Before merge/release: accept the documented custom-calculator performance tradeoff, retain durable provider test registrations, and resolve the RabbitMQ sample dependency in its own release work. Redis results do not establish non-Redis runtime coverage. The core sample advisory correction is included. Existing merge/release holds remain.

Verification and implementation details

Published head: 143f24fe11ebfc1be0f17cecea2df8c17353b14e; base: 3907f54ab0d70a6acd0073c5121abd4f6358afcc. Fresh main be10785cfed4e1850350dfaa024e1c3f25b7fac7 is included.

The latest regression reproduced retained memory in all nine new cases before the fix. After routing expired preparations through UpdateEntry, all 18 focused cases passed, including preservation of live stored entries. The integrated full suite ran 2,468 cases: 2,446 passed and 22 existing skips. That earlier local build retained Aspire and sample MessagePack warnings; the later sample-only correction pins patched MessagePack 2.5.303 and passed current hosted CI. The ordinary push workflow's CI-package publication step succeeded; release-package publication was skipped.

Existing custom calculators still receive concrete dictionaries. The read-only overload/option is additive and requires support for ordinary dictionaries and immutable read-only views. Oversized fallback estimates obey MaxEntrySize. Cloned reads preserve incremental list writes; uncloned raw reads retain existing shared-value behavior. Sparse storage and lazy key/value enumeration have allocation regressions in the core suite.

Historical paired measurements at 1,000 items with cloning disabled showed 3.27× time and 37× allocation versus main on the compatible custom-calculator path. These timings were not rerun at this head; the read-only opt-in avoids the dictionary copy but does not eliminate immutable-update allocation. This is an explicit compatibility/performance decision, not a universal performance claim.

Combined validation on 2026-10-23 used isolated core 70533664e7dfccf011fae4ddb5ef41c95dc68d63, containing #570 d2c24efd, #580 53ce936b, #582 3907f54a, #586 fffdf32f and this head. Its source tree is identical to the prepared combined candidate; only ancestry changed. No synthetic branch was pushed.

• Full core: 2,506 cases, 2,484 passed, 22 existing skips, zero failures/errors.
• Microsoft APICompat 10.0.402 against main be10785c: Foundatio net8/net10 and TestHarness net10 passed without suppressions.
• Refreshed sibling source build: Foundatio.All.slnx with ReferenceFoundatioSource=true passed, two existing Aspire warnings. RabbitMQ main 340357f1 required an isolated AppHost MessagePack 2.5.303 pin; no audit suppression or provider publication occurred.
• Redis 8.10.2: 796 discovered cases, 778 passed, 18 existing skips, zero failures/errors. All 30 lock cases and all 12 additional direct/scoped cache cases explicitly passed. Provider-output core/TestHarness DLL hashes exactly matched the combined build.
• Redis test checkout combined main 06668733, registration PR Bump Microsoft.Extensions.Hosting from 3.1.1 to 3.1.2 #191 ae79c527, and six shared cache-contract Fact overrides in each direct/scoped fixture. These test-only registrations are unpublished; provider-main discovery must not be inferred from the isolated run.

Tests used a dedicated loopback-only Redis container, removed afterward. No non-Redis runtime coverage or new timing benchmarks are claimed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-23T13:54:16.375212Z	ca9b5c9	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f13dca9648

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".