# Chapter 9 — registration & DSL transcripts (captured)

Real output from the verification machine (same environment as the earlier transcripts). The Chapter 9 overlay (ch09-registration) carries the registration wiring — modified copies of four repository files at their true paths, plus the recipe and golden fixtures — applied into a scratch clone; the main repository is never modified.

# 1. The wiring, as a diff summary

Four files change, each by the smallest honest edit, each following the shipped pattern for the eight existing paradigms:

• lexer.go — "critique": {} joins the reserved keyword set (between review and revise, its declaration-order neighbors).
• parser.go — critique joins the directive-item case list (the generic parseDirectiveItem handles the grammar).
• paradigm_options.go — criticOptions lowers the directive payloads (mirroring reflectionOptions line for line) and newCriticReviseBody runs authored body steps in-process via ExecuteDelegateCore.
• step_core.go — case "critic": in buildAgent’s switch (mirroring the reflection case; the default branch’s ErrContractViolation guard is untouched).

# 2. Goldens: untouched, then one deliberate baseline

With the wiring applied, every existing golden passes byte-identical, and only the new recipe lacks its triple:

--- FAIL: TestGoldenTraces/11_critic
    golden_trace_test.go:49: read golden testdata/golden/11_critic.trace.json:
    no such file or directory (set UPDATE_GOLDEN=1 to create)

Deliberate re-baseline, then the byte-identical proof:

$ UPDATE_GOLDEN=1 go test ./named/euclo/thoughtrecipes/ -run TestGolden
ok  	codeburg.org/lexbit/relurpify/named/euclo/thoughtrecipes	0.270s
$ go test ./named/euclo/thoughtrecipes/ -run TestGolden -v
--- PASS: TestGoldenPlans (0.00s)
--- PASS: TestGoldenGraphs (…)      [all existing triples byte-identical]
--- PASS: TestGoldenTraces (0.13s)
ok  	codeburg.org/lexbit/relurpify/named/euclo/thoughtrecipes	0.273s

The new 11_critic triple records the plan (recipe lowered: agents fixer/reviewer, the critic step), the graph (valid, 2 nodes), and the trace — whose exec error is the governance bar reaching the DSL harness:

{
  "graph_valid": true,
  "start_node": "run.10.5.reviewer.0.execute",
  "exec_success": false,
  "exec_error": "critic agent requires a permission checker (fail closed)"
}

The golden harness’s minimal deps carry no permission checker (the same honesty that records react agent missing language model in the shipped traces) — and the critic’s fail-closed refusal lands in the same protocol. The bar survives the DSL path; that refusal is the wiring proof.

# 3. Lint parity, both directions

A workspace carrying the critic recipe lints clean:

$ relurplint --check recipes --workspace ~/code/stringkit
relurplint: all checks passed (0 diagnostics)

And a D1 violation — a revise block with no critique before it — is rejected by the linter with the loader’s own error, at coordinates:

relurpify_cfg/euclo/critic_broken.erpe: [error recipes.contract]
  …/critic_broken.erpe:11:3: directive "revise" requires directive
  "critique" in the critic paradigm contract

# 4. The pinned offline run

relurpish chat --offline --workspace ~/code/stringkit, with critic_review.erpe materialized into the workspace, instruction: “explain and critique the slugify function in stringkit.py”. The deterministic route selects the critic recipe (telemetry route record):

family_id:investigation  recipe_id:euclo.thoughtrecipe.critic_review
route_kind:capability    name:euclo.thoughtrecipe.critic_review

One HITL approval banner (kind=execution | risk=medium), approved for the session ([s]), then the graph runs. The run completes with the loop-close honesty guard firing — the same flag Chapter 1 met — because the offline backend’s scripted responses are not valid verdict JSON: the critic’s strict decode fails, the D6 retry fires exactly once, and the failure protocol records the outcome. Telemetry carries the receipts (node euclo.execute_thoughtrecipe, the route record naming critic_review, critic as the executing paradigm).

The honest reading: addressability is proven end to end — load, lint, route, approve, execute, record. Critique quality under the offline backend is not; a scripted model that cannot emit verdict JSON cannot pass strict decoding, and the receipts say so. Real-model behavior is Chapter 12’s subject, and the runner (critic-runner) remains the deterministic behavioral proof.

[NEEDS CONFIRMATION: recipe dispatch selection rules — the first instruction ("critique …" with no investigation-family keyword) routed to the default recipe; the dispatcher's exact keyword-vs-family mechanics get pinned during Chapter 10 research.