[STRML]

Closes #167.

AUTODREAM_L1_REQUIRE_SERVED=1 decided a cached record was the runner's own by reading its keys, which a worker can write too. And every findings check used jq -e, which judges only the last value in a file.

Shape a substituted model writes	Before	After
{"error":"unreadable","findings":[]}	kept as the runner's, verdict skipped, session never retriaged	no runner marker, held to its verdict, retriaged
{"no_engine":true,"error":"x","findings":[]}, source has no engine	kept	dropped, the runner writes its own marked record
{"findings":[forged]} then {"skipped":true,"findings":[]}	passed every check, L2 read the forged finding	not a result, gate or no gate, retriaged

runner writes a record ─▶ <hash>.json + <hash>.json.served = "runner"
                                              │
drop_unacceptable_l1_cache ─▶ verdict runner + findings [] ? keep : verdict match ? keep : drop

• The dispatcher writes .served reading runner beside each of its five records (unreadable, noise gate, normalize failure, no engine, final-round stub). That marker is the only exemption from the verdict.
• A no-engine record is still dropped by its shape once its source resolves an engine. A forged one buys a retriage, nothing more.
• L1_FINDINGS_OK under jq -s (one value, an object, a findings array) replaces the five jq -e ".findings | arrays" checks. It is the rule merge-chunks.sh --check already applied to chunk answers.
• Runner records written before this change have no marker, so with the gate on they are triaged once more and come back marked. With the gate off nothing changes beyond the single-value check.

Tests: test_require_served_trusts_only_the_runners_own_marker and test_a_findings_json_with_two_values_is_not_a_result, 13 assertions, all red on main. Full suite 3,097 passed. The 4 failures only happen inside the Claude Code sandbox (no ps, no TTY) and are the same as on main. Both shellcheck gates are clean.