feat(runtime): integrate semantic systems and deterministic replay

[PhilipJohnBasile]

Summary

This PR integrates two native MTPLX system families on the clean 2.9.1 runtime baseline.

Semantic runtime intelligence

• exact semantic SessionBank anchors derived from verified chat-template prefixes
• a safe-point memory governor that mutates cache budgets only while holding the model lock and outside active foreground, scheduler, restore, commit, MTP, and postcommit work
• opt-in expert-locality instrumentation that measures MoE working sets and reuse without changing router outputs

Deterministic replay and evaluation

• privacy-safe request capture with prompt and response content disabled by default
• deterministic counterfactual candidate replay and evaluator isolation
• bounded candidate/evaluator timeout handling that returns without waiting for timed-out worker callbacks
• regression and promotion gates with explicit, non-automatic promotion
• trace-parity comparison for locating the first divergent tensor boundary
• bounded request-capture bookkeeping that removes completed, missing, and pruned path mappings

Operations surface

• GET /v1/mtplx/systems
• a Systems dashboard tab showing installed, wired, enabled, sampled, skipped, and inactive states
• the previously reviewed Native dashboard command surface
• dashboard compilation enforced before wheel and release packaging
• a permanent native-systems macOS CI job

Safety and scope

• no external runtime dependency
• no automatic candidate promotion
• no router mutation from locality telemetry
• no live cache mutation outside proven safe points
• no prompt/response body persistence unless explicitly enabled
• nested credential-bearing fields are recursively redacted

Audit hardening

The reviewed integration head 2f3d13996fe0b9eb8b519eb13ae118a8a4983eee received a focused audit pass covering timeout behavior, capture-map lifecycle, memory-governor synchronization, dashboard packaging, and permanent regression coverage. Temporary patch/materialization files were removed from the published branch.

Final validation

Exact head: a9ce9556c07bf495de31d121c03ce58342149fc4

• 122 focused semantic-anchor, memory-governor, expert-locality, deterministic-replay, trace-parity, request-capture, runtime-systems, and dashboard-endpoint tests: PASS
• 819 compatibility tests across public CLI, runtime KPIs, OpenAI server/bridge, thermal, cache-state, and no-MLX-import contracts: PASS
• Python compilation and Git whitespace checks: PASS
• production TypeScript/Vite dashboard build: PASS
• build / dashboard + wheel + fresh-venv smoke: PASS
• repository hygiene: PASS
• ci / no-mlx-smoke: PASS
• ci / native-systems: PASS

The PR is intended for squash merge into clean/upstream-2.9.1. Promotion from that integration baseline to main remains a separate reconciliation step.

[github-actions]

Resolved

This historical staging failure was fixed. Ruff now normalizes the touched test import block and validates only audit-introduced lines in legacy files.

Final audited PR head: a9ce9556c07bf495de31d121c03ce58342149fc4

• build: PASS
• hygiene: PASS
• ci / no-mlx-smoke: PASS
• ci / native-systems: PASS

[github-actions]

Resolved

The runtime, tests, and dashboard had already passed in this historical staging run. The remaining publication failure was removed by separating the audited runtime commit from authenticated workflow-file updates.

Final audited PR head: a9ce9556c07bf495de31d121c03ce58342149fc4

• focused native-systems tests: PASS
• compatibility tests: PASS
• production dashboard build: PASS
• wheel/fresh-venv build: PASS
• permanent native-systems CI gate: PASS

[PhilipJohnBasile]

Final merge receipt

PR #16 was squash-merged into clean/upstream-2.9.1 as 88954d81dad0c40e02df4b865b838faf1b14f750.

Final reviewed head a9ce9556c07bf495de31d121c03ce58342149fc4 passed:

• dashboard + wheel build and fresh-venv smoke
• repository hygiene
• compatibility/no-MLX CI
• permanent native-systems CI
• 122 focused native-systems tests
• 819 compatibility tests
• production TypeScript/Vite dashboard build

The two historical materialization-failure comments are marked resolved, and temporary audit PR #17 is closed.

[PhilipJohnBasile]

Mainline promotion and upstream handoff

The audited squash commit 88954d81dad0c40e02df4b865b838faf1b14f750 is now the fork's default main head. The former main is preserved exactly at archive/main-pre-2.9.1-20260823 (f440843729ac3287de8de0fc05bf5855b10d93fd).

The clean one-commit contribution has been opened against the original MTPLX repository as upstream PR youssofal#336. The upstream PR is mergeable, requests review from @youssofal, and runs upstream build, hygiene, compatibility, and native-systems CI.