[joelteply]

Split the 9,823-line llama-server owner into lifecycle, launch preparation, readiness, paging and reconciliation modules. This advances the Serving module's build-loop work: each responsibility can be reviewed separately while public paths and serving decisions remain unchanged.

The root is 145 lines, production children at most 608, tests at most 969. All 2,862 original full comment lines and all 62 test annotations are retained. Logging targets and subprocess selectors follow their original contracts. No new test setup or duplicate implementation is introduced.

Rebased onto canary 8e26442 after #5239. Independent source review transferred APPROVE to e887947: serving source and probe macros are byte-identical to reviewed dbc6064 (304 function definitions, public paths, logging and test selectors). Source-hygiene passes at this head with the directly measured baseline of 36,308 excess lines across 10 files. Windows installer passes; full core tests and Windows compilation remain pending. No serving-node compilation or deployment and no measured latency or learning claim.

If another split lands first, remeasure the baseline on the merged tree before integration.

[joelteply]

Validation at ef44115: Linux job118903846248 passes with 75 warnings, equal to baseline. Main suite8943 passed/47 ignored/42 filtered; explicit moved engine-install subprocess executes one test, retirement fixture and teacher custody fixture pass. Additional Postgres11 and doc guards2 pass. Binding check passes. Windows core artifact/custody and integration type-check steps pass; durable-stop regression still running. This is test evidence only, not deployment, latency improvement or learning proof. Hygiene remains blocked on the previously reported scanner floor; no baseline increases.

[joelteply]

Independent source review — Continuum PR5233

Verdict: source APPROVE exact dbc6064 versus 437ee5f. CI remains the compilation/platform gate; no merge approval before exact-head checks are green.

Read AGENTS.md, CLAUDE.md plan, README and changed module facade/macro/ratchet. This serves persistent citizens by making serving lifecycle code independently reviewable and advances phase 1 / goal 4 (No monolith), under Codex serving/inference ownership. No builds or owner checkout edits performed; checkout was clean at the reviewed head.

Independent mechanical/source checks:

• Compared old monolith against all resulting module/test files using normalized line multisets and a tokenized function-body comparison. 304 function definitions on each side. After only normalizing inserted tracing target arguments and relocation-qualified inference sibling paths, the only remaining function differences were two subprocess test-selector strings (correctly updated to tests::engine_install and tests::retirement) and indentation in an escaped-newline assertion string (same Rust string value). This is a source consistency check, not a compiler or full Rust parser claim.
• 62 test annotations on both sides. No original full comment line missing. Longest resulting llama-server file is tests/retirement.rs at 969 lines; all resulting files are below 1,000.
• Parent facade reexports public serving types/functions/constants at the original llama_server path. Child helper visibility is widened only to pub(super) where sibling impl/test responsibilities require it; existing crate-visible exports remain explicit. LlamaServerProcess inherent implementations remain on the same public type.
• Inference sibling references formerly using super are qualified through crate::inference at the deeper module depth. No stale include_str/include_bytes source-contract paths referencing the split monolith found in the searched core/tests/scripts surfaces; module-relative subprocess test names were adjusted.
• Every tracing/probe invocation in the moved module tree explicitly retains continuum_core::inference::llama_server as target. New probe! target arm retains INFO and probe_class assignment identically to the existing arm; old macro syntax is untouched. Macro compilation smoke invocation added.
• The hygiene baseline change is a downward ratchet (54,053 to 45,230), not a waiver or increased allowance.

No concrete behavior regression found in this exact split. Does not establish live serving behavior or learning, and does not close the separate AIRC/Windows incidents. Exact-head hosted CI and the required live validation belong to integration/deployment owners.

[joelteply]

Landed via #5245 (merge-train, canary 741c118): this PR's own commits, cherry-picked unchanged. Its review stands at the head named in #5245's body.