Record role tool sets and MCP memory gates - #910

Scope

Role sets now cover their observed necessary servers with client-specific selection. Exec readers require EVALUATION, the narrowest existing contract covering all nine observed servers; conflicts with live lane exclusions go to the CC and intentional exclusions require a deciding file and decision. The record separates historical trial sets from live settings and aligns every named catalog disposition with its actual verdict row.

• Base commit: 72e1f943d951e6aff4200729db88121a8d38d7da.
• Previous reviewed head: 897b87229c2d59f93244ae6e4b9038de4a739469.
• FIX2 head: 779fa242b4b3dd9c25f53295de49b3d0e5b2ab7d; one additional repair commit and one push, no rebase.
• Lane: lane:foundation.
• PR paths: docs/role-tool-sets.md, docs/role-tool-sets.json, docs/convergence-architecture.md, adoption/README.md, and documentation metadata in manifests/evidence.json. The final FIX2 delta updates only the two role records; earlier architecture/adoption links and manifest integrity remain part of the five-file PR.
• Ten role rows, seven historical process contracts, seventeen tool records, thirteen pending/source-context slots.

The live Codex exclusions are bound per named lane to role-mcp-off.json v2 SHA256 77c91cdc0433f54ecacb070a4581ec7f20443d9d5835e7fb938f3994e6563107. CORE is explicitly the historical trial base; all four trial arms used role_mcp_off=skip. Claude's twelve-component strict selection contract needs grafana and agentsview additions for the CC/co-op; all observed necessary servers remain covered by role-specific contracts. Unknown coverage cannot pass a server reduction or shared-MCP topology decision.

Stage 1 uses valid array-index catalog verdict pointers: Context Mode and jCodeMunch are measured trade-offs, Headroom and GitHub CLI candidate:cli-cli are winners for their respective jobs, CBM and agentsview are conditional, and Hindsight is unqualified. Semble/browser links are layer context only. The coordination composition is pending and owned by orch-records. Hcom's tree and installed pins are distinguished, and grafana-observation pins the MCP connector. The architecture paragraph is appended to restore the test-enforced citation at line 134. Optional QMD cohort numbers are omitted because original observer files are not bound to this numerical packet. Each historical arm separately cites a hashed release.out:2 record for role_mcp_off=skip. The agentsview table is contiguous, api-actions has review-specific acceptance, and the Claude schema source identifies the 1000-byte span, start and term offsets. The profile readback came from the identity-launcher binary reporting Codex 0.161.0; launch metadata and native execution stay separately scoped.

SOTA sources

• openai/codex 0.161.0 @979011409de0a60b52f179721948e65531d26144: MCP enabled filtering and bounded child-role fields; installed CLI help and official profile reference. Historical trial execution used 0.162.0 and remains separate.
• Installed Claude Code 2.1.295 help and embedded MCP schema: strict MCP selection and optional alwaysLoad prompt deferral; inspected schema span SHA256 98c935556619bbd70ec55ddd7b923bb1b8ff797de883af82e9af1401f47e1cc9. Native child MCP scopes remain separate from independent-process selection.
• aannoo/hcom 0.7.27 @2c5f343b2f9ec4bf2acf49c0431860e7c2ae578b is the retained tree pin; installed 0.7.28 is recorded separately at b2a7c192003e7fd67ed93265289e4ac36276f965.
• grafana/mcp-grafana @7514746a763b794ead31902b7821bf73280503b4 identifies the connector; the Grafana server is a separate dependency. kenn-io/agentsview 0.44.0 @413a87f7bfbd67b2815b1119ac51abc1efbeeaba supplies the existing session archive route.
• oraios/serena @c6fbd1c5932df2494ffa0020af5a9fbe80b82143: README.md:226–245 and cli.py:275,369; akitaonrails/ai-memory 2.4.1 @433a19f3d54dea287571b1423591db2a89965fa9: cookbook:793,1518 and tagged release assets; rtk-ai/rtk 0.51.0 @e001f773f80b22b7dc4c7a79521b30e35aaef026: README.md:103,129 and awareness contract.
• Retained CBM/QMD/Headroom and other vendor transport sources remain pinned in the JSON. QMD's dated pilot is historical context only. Current catalog verdicts, orchestration composition and bounded AO adoption are the named deciding records. The remaining corrected verdicts are foundation.json#/layers/4/alternatives/2 (jCodeMunch), /layers/19/winners/1 (GitHub CLI), and /layers/15/alternatives/4 (agentsview). The GitHub CLI winner establishes source-reviewed GitHub operations; generic git and PR lifecycle acceptance remain separate.
• The installed ECC search-first skill supplied the coordinator preflight; context-mode supplied bounded source processing. Existing pinned work and trials were reused.

Evidence-class table

Each test module ran under nice/ionice with timeout 600. Counts above exclude skips from passed totals. Complete returned stdout/stderr and return codes are retained in the private verification record. These are existing repository checks, not vendor acceptance.

Per-tool PSS, fresh concurrent sharing, effective Claude registration readback, generic role/phase mappings and role 99's native client remain unmeasured. No runtime configuration, model route, provider trial or deployment changed. FIX2 incorporates the CC micro's 2 P2/7 P3 and the designated GPT micro's same table P3. The read-only completeness critic reports PASS WITH BOUNDARIES. Both designated exact-head reads, CI, required pre-cue tooling, the CC cue and owner decisions remain required for landing.