conformance-gate¶

A pending design contract — the scope, not the implementation. It marries the two halves the tree already carries: ci-gate (the non-bypassable CI backstop running spex lint + tsc) and spec-forge (the read-only forge tracer that resolves an issue/PR to the node it serves). The gap between them is the whole node:

Not a standalone feature — one direction of the deliver-port framework. An external PR arrives as a session-origin pr-origin session defaulting to destination: verdict; delivering that verdict (a Check + a sticky comment) is what this subtree builds. The cluster is the deliver port's verdict driver (plus the forge driver's write verbs it stands on) — the inward half (external PR → reviewed → accepted or rejected), mirror of the outward half (our own session → a PR). When a maintainer accepts, the same session simply retargets to destination: trunk — no separate merge path.

take the session-eval derived-evidence model — today fed only by an internal session's worktree — generalize its root to any PR branch, add an agentic conformance verdict on top, and write that verdict back to the forge's execution plane (a Check + a sticky comment).

This is the LLM judge spec-lint explicitly deferred: lint keeps the spec↔code graph honest in the commit path; whether the code still honors what the spec says is judged here, async, on the forge.

the pipeline this parent scopes — determinism first, agency only where meaning must be judged¶

A forge event (PR opened/synchronized) runs spex forge gate <PR> in CI. Four layers:

Tier 0 — deterministic, blocking. Reuse ci-gate/manager-cockpit's gates: spex lint errors
tsc --noEmit + conflictsWithMain. Red here fails before any agent runs — cheap, no LLM.
Tier 0.5 — deterministic, signal. The spec-touch matrix: route each merge-base-diff file to its node via the code: graph, classify (spec touched?, code touched?) per node. The load-bearing cell is code changed / spec untouched — the silent divergence the dogfood forbids.
Tier 1 — agentic, the verdict. One judge per touched node, structured output. Owned by conformance-judge.
Tier 2 — agentic, optional (default off). On diverges/spec-missing, propose the spec edit or missing eval scenario as a forge suggestion — never an auto-push or auto-merge.

the load-bearing decision — writing back does NOT break the read-only contract¶

spec-forge's rule is "never write a node's git-derived status/version". A Check Run and a PR comment are the forge's execution plane, not the spec graph. The verdict never sets a node's version or status; a merge still flows back only through git. So the write-seam is a new capability axis, not a violation — definition (the graph) and execution (the forge) stay un-crossed, exactly as the port requires.

the decomposition — what each child owns¶

forge-write-seam — the forge driver's write verbs the read-only port lacks (open/update a PR; post a Check; upsert a sticky comment), behind the same host-agnostic seam — what both the outward (session → PR) and inward (Check + comment) deliveries go through.
conformance-judge — the verdict driver's core: the per-node agentic intent verdict (Tier 1) and its deterministic Tier-0.5 spec-touch input.
forge-gate — the orchestration: spex forge gate <PR> + the CI workflow that runs it; also the one change generalizing the session-eval evidence root from a session worktree to any PR ref.
dashboard-prs — the dashboard surface: PR badge, review lane, PR proof overlay (the open-PR sibling dashboard-issues deferred).

scope¶

The whole subtree is pending — design only, no code. Out of scope until a build decision: any implementation, a second (GitLab) driver, auto-remediation past a Tier-2 suggestion. This parent owns only the cross-cutting shape and the load-bearing decision; ci-gate keeps the deterministic backstop, session-eval the evidence engine, spec-lint the graph rules.