fix(codex): reuse shared recall compressor

[fancyboi999]

Description

Codex auto-recall maintained a private compression path instead of using the repository's shared recall-compression core. It therefore launched nested codex exec compression for every non-empty result, including bounded context, and could not reuse an identical digest.

This change routes Codex's injection-ready context through compressRecallContext(). The Codex adapter continues to own model/profile selection, process launch, timeouts, provider overrides, and runtime_failed recovery; the shared core owns admission, cache identity, prompt/output validation, and served-URI repair.

The resulting behavior is:

• context below recallCompressMinInputChars (default 1500) is injected without a nested model call;
• an identical query/context/URI set reuses the Codex-local digest cache;
• eligible cache misses lazily resolve the configured Codex compressor profile and launch codex exec;
• compressor failures preserve the existing deterministic fallback, including the legacy /recall path fixed by fix(codex): compress type-quota recall results #3248;
• recallCompressMinInputChars=0 preserves always-compress behavior.

Human Involvement

• A human participated in the implementation or review loop
• This PR was generated entirely by AI agents without human participation in the loop

Related Issue

Fixes #4444

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Performance improvement
• Test update

Changes Made

• Replaced the Codex-only compression prompt/normalizer with the generated shared compression core while preserving the Codex profile and subprocess adapter.
• Added the persistent OPENVIKING_RECALL_COMPRESS_MIN_INPUT_CHARS / plugin.codex.recallCompressMinInputChars setting and documented the default/cache behavior.
• Hardened the shared owner so header-only cached digests and repair results with no surviving cited bullet fail open to caller fallback instead of injecting an empty digest.
• Preserved served URIs containing spaces and decoded escaped citations before provenance comparison; synchronized the generated shared-module copies.
• Added production-entrypoint regressions for short-input admission, repeated-request cache reuse, forced compression, provider override, NO_RELEVANT_MEMORY, and failure fallback.

Root Cause

#3248 restored compression on the Codex endpoint success path, but its item-based compressor predated the shared core. #3534 later introduced recall-compress-core.mjs and wired Claude Code through it; the generated file was copied into the Codex bundle without being imported by Codex auto-recall.mjs.

The first real smoke through the shared path also exposed two shared validation gaps: a fully rejected URI-repair result retained the digest header, and the URI scanner truncated legal served paths containing spaces. Both fixes live in the shared owner rather than adding Codex-specific exceptions.

Compatibility and Trade-offs

No server, REST, MCP, identity, or persisted-memory format changes. Small recalled blocks are now injected directly by default instead of paying a second model call for semantic filtering; the server context assembler has already applied retrieval, ranking, budgeting, and cross-turn dedup. Operators that prefer model filtering for every non-empty result can set the minimum to 0.

The digest cache lives under OPENVIKING_CODEX_STATE_DIR, separate from Claude Code state. Legacy servers and compressor failures retain bounded deterministic injection.

Testing

• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have tested this on the following platforms:
  • Linux
  • macOS
  • Windows

Commands and results:

• Baseline reproduction on current main: Codex auto-recall test file 10 passed, 2 failed (short: 1 != 0 nested calls; repeat: 2 != 1).
• npx --yes node@24 --test examples/codex-memory-plugin/scripts/auto-recall.test.mjs examples/memory-plugin-shared/recall-compress-core.test.mjs examples/memory-plugin-shared/sync.test.mjs: 20 passed, 0 failed.
• Full memory-plugin matrix from .github/workflows/pr.yml on Node 24: 374 passed, 0 failed, 1 skipped out of 375. The skipped test is the existing environment-gated live DSH recall test.
• All six installer scripts checked by the memory-plugin CI job pass bash -n.
• git diff --check: passed.
• Real macOS smoke against OpenViking 0.4.16 with a configured gpt-5.6-luna/low compressor: first eligible request produced a valid cited digest; the identical second request completed in about one second with no compressor profile resolution, confirming a cache hit. No Spark request was used for this verification.

Checklist

• My code follows the project's coding style
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Screenshots (if applicable)

Not applicable; this changes hook control flow and generated context.

Additional Notes

• feat(retrieval): add context-assembler admission policy #3687's optional server admission policy and perf(codex): skip recall for control-only prompts #3881's control-prompt short circuit are complementary; neither supplies the client minimum-input gate or digest cache.
• Windows was covered only by the existing mocked spawn tests; no real Windows Codex/OpenViking smoke was run.