perf(cognition): split files >1000 LOC — evaluator.rs (1231), admission.rs (1225), model_resolver.rs (1232)

[joelteply]

airc-queue card

Coordinates work via the AIRC queue substrate (airc#562). Edit this card by commenting OR by running airc queue claim/airc queue release/airc queue heartbeat (later PRs).

{
  "kind": "airc-queue-card-v1",
  "id": "#1208",
  "owner": "claude-tab-1",
  "status": "merged",
  "evidence": "Adopted existing GitHub issue into airc queue.",
  "next_action": "Triage, claim, or close this adopted backlog card.",
  "last_heartbeat": "2026-05-14T20:37Z @ 2405c83c8"
}

Close this issue when the work is done (status=merged/abandoned).

Original issue body

Pre-adoption body

Smell

Three files in cognition/persona Rust modules exceed the 500 LOC lint cap (likely cap; verify against the actual rust-max-lines config). Each is a sign of missing module split:

• `workers/continuum-core/src/persona/evaluator.rs` (1231 LOC) — mixes 5 admission gates + sleep mode + rate limiter + social signals
• `workers/continuum-core/src/persona/admission.rs` (1225 LOC) — core state machine + semantic-similarity scoring + context hashing
• `workers/continuum-core/src/cognition/model_resolver.rs` (1232 LOC) — resolver logic + tier matching + silicon residency filtering

Fix

For each, extract concerns into submodules:

evaluator.rs split

• `evaluator/mod.rs` (orchestration only, <500 LOC)
• `evaluator/sleep_mode.rs`
• `evaluator/rate_limiter.rs`
• `evaluator/social_signals.rs`
• `evaluator/gates/{social,topic,addressed,recent_self,no_op}.rs`

admission.rs split

• `admission/mod.rs` (state machine, <500 LOC)
• `admission/scoring.rs` (semantic-similarity)
• `admission/context_hash.rs` (hashing)

model_resolver.rs split

• `model_resolver/mod.rs` (resolver, <500 LOC)
• `model_resolver/hw_tier_matching.rs`
• `model_resolver/silicon_residency.rs`

Same pattern as the `MetalMonitor` + `FootprintRegistry` modular submodule splits already shipped (tasks #45 + #46 in local task history).

Acceptance

• All three files under 500 LOC after split
• No public API surface change
• Tests still pass

Refs

• Parent: architecture: Rust AI bus scheduler for GPU-resident throughput, memory paging, and kernel offload #1201, RTOS-for-AI-Throughput: Rust-owned memory paging + GPU-native inference (umbrella) #1203
• Prior art: Build(deps-dev): Bump @typescript-eslint/parser from 8.29.1 to 8.31.1 #45 (MetalMonitor split), Build(deps): Bump glob from 11.0.1 to 11.0.2 #46 (FootprintRegistry split)
• Source: Rust cognition smell-hunt survey 2026-05-14

Status log

• 2026-05-15T14:48Z — merged via PR refactor(persona): split admission.rs (1225 LOC) into mod + recipes (#1208) #1251 @ cc885c6 (closed by github-actions[continuum#1142])

• 2026-05-14T22:35Z — merged via PR refactor(cognition): split model_resolver.rs (1232 LOC) into mod + types (#1208) #1249 @ 0ca0f57 (closed by github-actions[continuum#1142])

• 2026-05-14T21:28Z — merged via PR refactor(persona): split evaluator.rs (1231 LOC) into focused submodules (#1208) #1242 @ f9d080b (closed by github-actions[continuum#1142])

• 2026-05-14T21:28Z — airc-8a5e-bc43 -> status=merged

• 2026-05-14T20:37Z — claim by claude-tab-1 -> status=review

• 2026-05-14T20:26Z — claim by claude-tab-1 -> status=in-progress

[joelteply]

Cross-link: I filed #1205 covering the modules/cognition.rs IPC dispatch monolith (separate layer — 1390 LOC, 37 arms in one match). Pair them — your card splits persona/* internals, #1205 splits the IPC routing. Same compression-principle move; same sprint. — claude-tab-1

[joelteply]

Resolved the evaluator slice via PR #1242, merged to canary at f9d080b.\n\nReview validation: public re-exports stayed stable at persona::evaluator/persona, extracted modules preserve the old sleep/rate/adequacy logic, and local validation passed: cargo test -p continuum-core --features metal,accelerate persona::evaluator --lib (32 passed).\n\nFollow-up decomposition for admission.rs/model_resolver.rs should continue as separate focused queue cards rather than keeping this broad card open.