eval: cover Agent Spec launch method selection

.gitignore

@@ -3,6 +3,7 @@
 .sandbox/
 .personas/
 .eval-runs/
+tmp/
 
 # noise
 .DS_Store

CATALOG.md

@@ -1,7 +1,7 @@
 <!-- Generated by bin/generate-catalog.sh from cell KDL and evidence/*.tsv. Do not edit. -->
 # Current eval catalog and evidence
 
-This is the sole current corpus surface: **59 included cells** (19 model-free,
+This is the sole current corpus surface: **60 included cells** (20 model-free,
 40 model-backed) and **4 retired exclusions**. Inventory fields are derived from the active
 KDL and executable judge scripts. Latest accepted PASS evidence remains distinct from the append-only run
 history: a failed last run is visible without being advertised as accepted evidence, and cells with no
@@ -42,6 +42,7 @@ harness-native loader plus canonical hook file.
 | `host-lock-health-negatives` | model-free | — | 0 | none | `30s` | 5 | — | **NO STRUCTURED RUN** |
 | `inbox-hygiene` | Claude | `claude-sonnet-5` / medium | 1 | low | `900s` | 4 | — | **NO STRUCTURED RUN** |
 | `incident-response` | Claude | `claude-sonnet-5` / medium | 2 | medium | `1200s` | 5 | — | **NO STRUCTURED RUN** |
+| `launch-method-selection` | model-free | — | 0 | none | `120s` | 5 | — | **NO STRUCTURED RUN** |
 | `license-mit` | mixed | `claude-sonnet-5+gpt-5.6-sol` / medium | 3 | high | `1200s` | 6 | — | **FAIL** 2026-07-28T13:27:18Z, 5/6, source [`b3cd5fb`](https://github.com/compoundingtech/evals/commit/b3cd5fbd98c11179a4555d0f9bbccfe98351a734), [receipt](evidence/stable-main-phase4-runs-20260728.json) |
 | `license-mit-codex` | Codex | `gpt-5.6-sol` / medium | 3 | high | `1200s` | 6 | **PASS** 2026-07-26, st2 `9d26245`, 1m39s, [`40ec23d`](https://github.com/compoundingtech/evals/commit/40ec23d3e068e5bd70b6d4b9a92a92d291868dce) | **NO STRUCTURED RUN** |
 | `managed-agent-color-env` | model-free | — | 0 | none | `90s` | 9 | — | **NO STRUCTURED RUN** |
@@ -102,6 +103,7 @@ while the last-run column makes a recorded failure distinct from a cell with no
 - `exec-lifecycle-logging` / `cell` — deterministic native exec lifecycle probe; no harness seat
 - `hook-integrity` / `cell` — deterministic native hook-installation probe; no live harness seat
 - `host-lock-health-negatives` / `cell` — deterministic native doctor negative probe; no harness seat
+- `launch-method-selection` / `cell` — deterministic proposed launch-method selection probe; no harness seat
 - `pty-attach-machine-stream` / `cell` — deterministic installed PTY attach composition probe; no harness seat
 - `managed-agent-color-env` / `cell` — deterministic managed-agent environment policy probe; no harness seat
 - `pty-attach-only` / `cell` — deterministic PTY dead-attach policy probe; no harness seat
@@ -116,7 +118,7 @@ while the last-run column makes a recorded failure distinct from a cell with no
 - `targeted-reconcile-isolation` / `cell` — deterministic selected reconcile isolation probe; no bus-connected agent
 - `two-networks-coexist` / `cell` — deterministic catalog-isolation probe; no harness seat
 
-These **20** rows are generated from structured exclusions. A cell-level row is required for
+These **21** rows are generated from structured exclusions. A cell-level row is required for
 every derived model-free cell. The one-shot docs cold reader is a model-backed grader, but not a bus-connected
 agent: it has no st2 identity, DING sidecar, or lifecycle hooks.
 

cells/launch-method-selection/README.md

@@ -0,0 +1,52 @@
+# launch-method-selection
+
+Contract-first, model-free acceptance coverage for the proposed start/resume launch-method shape discussed in
+[`compoundingtech/st2#124`](https://github.com/compoundingtech/st2/issues/124). The field names and defaults are
+not ratified. This cell therefore remains proposal-tracking evidence and does not update `AGENT-SPEC.md`.
+
+The four proposed cases are expected product reds while the current st2 grammar exposes one top-level `argv`.
+The legacy case is a positive control and must keep passing. Expected reds are classified evidence, never
+conformance passes.
+
+**Capabilities required:** `st2,pty,jq,awk,grep,sed,coreutils`. No model or provider. Every catalog, synthetic
+session record, workspace, durable marker, process, and PTY root lives below the eval-owned temporary catalog.
+
+## Closed-set cases
+
+| Case | Current result | Discriminator |
+| --- | --- | --- |
+| `explicit-start-new-session` | `RED/P01` | selects `start`, creates one new synthetic native ID, and preserves durable work |
+| `explicit-resume-exact` | `RED/P01` | appends parsed conversation records only to the declared ID, never a newer neighbor |
+| `unavailable-refuse` | `RED/P01` | refuses before either method launches or any session record changes |
+| `unavailable-start` | `RED/P01` | selects `start` only when that unavailable fallback is declared |
+| `legacy-single-argv` | `PASS` | preserves the existing single top-level `argv` launch behavior |
+
+`P01` means the proposed grammar is unavailable. Once a candidate parses it, behavioral failures are classified
+as `P02` wrong method, `P03` wrong session, `P04` wrong unavailable policy, `P05` lost durable continuity, or
+`P06` leaked eval-owned runtime state.
+
+The checked-in `AGENT-SPEC.md` remains pinned to an older portable runner and documents compact `command`.
+Current st2 declarations use one top-level `argv`; this proposal replaces that cardinality invariant with a
+successor that accepts either one legacy `argv` or launch methods, never an ambiguous mixture.
+
+## Evidence boundary
+
+The adapter writes synthetic JSONL conversation records and checks them by parsed `sessionId` and record type.
+It never infers continuity from file bytes or mtime. It does not call Claude or Codex and does not touch their
+homes or transcript stores. Provider-native facts remain the version-scoped evidence recorded on st2#124;
+this cell covers only provider-neutral selection and refusal.
+
+The candidate preserves the optional `resume.session` pin in its public Rust `agent-spec` model, but the
+current versioned `agents`/`tasks` CLI and launch receipts do not project it. This external behavior cell can
+therefore prove that the selected resume argv continues the exact adversarially chosen session, but it cannot
+independently prove parser-level pin preservation. The upstream model tests own that assertion until a public
+wire surface is deliberately versioned; this cell does not invent one.
+
+`judges/self-test.sh` plants missing-case, duplicate-case, false-promotion, unknown-gap, false-honesty, and
+cleanup-receipt mutations. The focused run remains red until every proposed case passes and the closed-set
+manifest is updated.
+
+```sh
+bash ./cells/launch-method-selection/judges/self-test.sh
+st2 eval ./cells/launch-method-selection/ --keep
+```

cells/launch-method-selection/fixture/adapter.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+mode="${1:?mode is required}"
+store="${2:?session store is required}"
+durable="${3:?durable marker is required}"
+receipt="${4:?launch receipt is required}"
+requested_session="${5:-}"
+
+test "$(<"$durable")" = "DURABLE-WORK-CONTINUES-8d31"
+mkdir -p "$store"
+
+fresh_uuid() {
+  seed="$(date +%s%N)-$$-$RANDOM-$store"
+  hex="$(printf '%s' "$seed" | sha256sum | cut -c1-32)"
+  printf '%s-%s-4%s-a%s-%s\n' \
+    "${hex:0:8}" "${hex:8:4}" "${hex:13:3}" "${hex:17:3}" "${hex:20:12}"
+}
+
+case "$mode" in
+  start)
+    session="$(fresh_uuid)"
+    printf '{"type":"conversation","sessionId":"%s","turn":"created"}\n' "$session" \
+      > "$store/$session.jsonl"
+    ;;
+  resume)
+    session="${requested_session:?resume session is required}"
+    test -f "$store/$session.jsonl"
+    printf '{"type":"conversation","sessionId":"%s","turn":"resumed"}\n' "$session" \
+      >> "$store/$session.jsonl"
+    ;;
+  legacy)
+    session="-"
+    ;;
+  *)
+    printf 'unknown adapter mode: %s\n' "$mode" >&2
+    exit 2
+    ;;
+esac
+
+printf '%s\t%s\t%s\t%s\n' "$mode" "$session" "$(<"$durable")" "$$" > "$receipt"
+exec tail -f /dev/null

cells/launch-method-selection/fixture/cases.tsv

@@ -0,0 +1,6 @@
+case_id	expected	gaps	contract
+explicit-start-new-session	RED	P01	explicit start creates a new synthetic native session while durable work remains intact
+explicit-resume-exact	RED	P01	explicit resume appends only to the declared session and never to a newer neighbor
+unavailable-refuse	RED	P01	an unavailable resume refuses without launching either method or mutating session state
+unavailable-start	RED	P01	an unavailable resume degrades to start only when that fallback is declared
+legacy-single-argv	PASS	-	the legacy single top-level argv form retains its existing launch behavior