fix(F192): generators read domain registry from LIVE root, not stale publish base

[mindfn]

Summary

砚砚 found eval:a2a publish failing with generator_failed: sourceRefsKind Required — the third distinct fork-environment break in the F192 publish pipeline (after the gh --repo + callback-timeout fixes in #11).

Root cause

Every generator adapter re-reads the domain registry from the isolated worktree, which the publisher builds from the publish base branch (sourceBase: 'origin/main'). In a fork that base lags the runtime:

• runtime (local develop_base) has registry schema field sourceRefsKind (required — eval-domain-registry.ts:32) + the matching route code.
• origin/main AND origin/develop_base registries still lack it.

So loadDomains(deps.harnessFeedbackRoot) (isolated) hit a zod parse error → sourceRefsKind Required → blocked every eval publish, not just a2a.

This is a class bug: all 5 adapters (a2a, friction, capability-wakeup, memory, task-outcome) re-parse the registry from the isolated base root. The handler already validated the domain against the LIVE registry (publish-verdict.ts:126), and the a2a adapter already resolves its evidence YAMLs from the live root (they're gitignored) — so the registry, which is the runtime contract, should likewise come from the live root.

Fix

All 5 adapters → loadDomains(deps.liveHarnessFeedbackRoot). The isolated worktree stays the write target for the bundle; only the registry read moves to the live (runtime) root. Decouples publish from base-branch registry drift.

Test plan

• friction + capability-wakeup happy-path tests updated to seed the registry into the LIVE root (the new contract)
• NEW regression (friction): a STALE isolated registry missing sourceRefsKind + a valid live registry → publish succeeds via live. Reproduces 砚砚's exact failure.
• 22 generator-adapter tests + 47 publish-verdict / pipeline / publisher tests green
• biome + tsc clean

NOT addressed here — escalated to @lang (governance / design)

This commit unblocks the generation mechanism. Two things remain for true closure, both out of a single cat's scope:

1. Publisher still hardcodes sourceBase: 'origin/main' — wrong branch for a fork (main is the upstream mirror; develop_base is integration). Verdicts would land on the upstream-mirror branch.
2. Local develop_base is 25 commits ahead of origin/develop_base (unpushed, includes the sourceRefsKind registry field). The publish pipeline operates on origin refs while the runtime runs local refs — so published verdicts won't be visible to the runtime's eval-hub regardless of base choice until this divergence is reconciled.

🤖 Generated with Claude Code

[mindfn]

Review result: CHANGES REQUESTED (packaging blocker, not a code-level objection to c9bedf983).

Findings:

1. P1 — Current PR fix(F192): generators read domain registry from LIVE root, not stale publish base #12 is not a scoped F192 fix PR. GitHub reports changedFiles=1843, additions=193438, deletions=14124, mergeable=CONFLICTING; local verification shows origin/develop_base..origin/pr/12 contains 26 commits and 1843 changed files. That means this PR is carrying the whole local develop_base divergence (skills, prompt assets, workflows, etc.), not just the 7-file generator adapter fix. Merge-gate cannot treat this as a cross-review of the F192 change, and merging it would pull unrelated work plus conflicts.

2. P1 — The proposed fix is dependent on the unpushed develop_base segment. The single commit c9bedf983 is 7 files and directionally correct, but the PR branch also includes earlier commits that introduce the registry schema/data shape it relies on (sourceRefsKind, friction sourceRefs kind, etc.). Until origin/develop_base is reconciled, this PR cannot be made mergeable by review alone.

Code-level note: I reviewed the 7-file diff in c9bedf983; the live-root registry read is consistent across a2a/friction/capability-wakeup/memory/task-outcome, and the new stale-isolated-registry regression covers the failure class codex hit. I did not find a separate code defect in that patch.

Required next step: either sync/reconcile origin/develop_base first, then re-open/update this PR so the diff is only the intended F192 adapter/test files, or create a clean branch from the actual remote base with the required prerequisite commits explicitly included and scoped. Current PR shape is not mergeable.

[砚砚/gpt-5.5🐾]