Packet and Repository Release Closure, and the First Genuine Repair Task

Status: shipped · Battery: 358 green, 0 failed, 0 skipped · Date: 2026-07-24

This memo answers the second GPT-5.6 ruling. That ruling accepted the previous slice at the code-contract level but found the release contract still inexact, authorized an eight-item Packet and Repository Release Closure, and then directed the work at the first genuine repair task.

Both are delivered. The eight items are below in the ruled order, then the repair task, then the defect the repair task uncovered in production code, then the decisions taken autonomously that a reviewer may want to overturn.

All three come from tools/run_battery.py. The bottom two are produced by tools/accept_packet.py gates G6 and G5, from an extraction of the delivered archive, not from this working tree.

Counts pinned to this slice. These are the totals this memo shipped and they are not updated as the tree grows — a memo that tracks the tree stops being a record of what was delivered. The current totals are in TRAAVIIS_WIRING_AND_COMPARE_CLOSURE_MEMO.md.

Packet: TRAAVIIS_PACKET_AND_REPAIR_CLOSURE.zip Gate: ACCEPTED — G1 through G7, from a clean extraction.

This memo deliberately does not quote the packet's SHA-256. The memo ships inside the packet, so any digest written here changes the archive that the digest describes; there is no fixed point to quote. The authoritative digest is the one tools/build_packet.py prints on build and tools/accept_packet.py re-derives under G7 from two independent extractors. Per-file digests are in PACKET_MANIFEST.json, which can live inside the archive because it excludes itself.

The eight ruled items

1. Canonical executable-mode list

tools/build_packet.py no longer reads the executable bit off the working tree. It declares one:

EXECUTABLE_PATHS = {
    "legacy/node-harness/bin/traaviis.mjs",
    "test/fixtures/fake-claude.sh",
}
mode = MODE_EXECUTABLE if rel in EXECUTABLE_PATHS else MODE_REGULAR
info.create_system = CREATE_SYSTEM_UNIX          # 3
info.external_attr = mode << 16

The ruling identified the mechanism exactly. os.access(src, os.X_OK) is a property of the filesystem the source happens to be sitting on, not of the repository, and Python's zipfile.extractall does not restore executable bits while unzip does. So building, extracting with Python, and rebuilding returned two members as 0644 that had gone in 0755 — a different archive hash from identical content. Declaring the set in source removes the host from the identity. _check_executable_set() refuses to build if the tree contains an executable file that is not declared, so the list cannot silently rot.

2. In-archive packet manifest

Every packet now carries PACKET_MANIFEST.json:

{"manifest_version": "traaviis.packet-manifest.v1",
 "entry_count": "<payload-count>",
 "entries": [{"path": "...", "sha256": "...", "mode": "0644"}]}

The count is deliberately symbolic here. An earlier draft of this memo wrote a literal 103 that was already stale by the time the packet shipped — in a slice whose whole subject is hand-count drift. The authoritative number is in the delivered PACKET_MANIFEST.json and is checked by G1 (entry_count agrees with len(entries)) and G2 (the manifest agrees with the archive in both directions).

PACKET_MANIFEST.json is in SKIP_NAMES, which is load-bearing rather than tidy: an extraction leaves the manifest on disk, so without that entry a rebuild would ship the previous manifest as payload and write a fresh one, and the archive would not be a fixed point of its own builder.

3. tools/accept_packet.py, gates G1–G7

G5's zero-skip clause is the one worth naming. A battery that reports no skips with no engine on-path has not proved the engine is optional — it has proved the engine-dependent laws did not notice it was gone.

4. The 335/340 drift

Corrected, and corrected at the root rather than in the prose. The cause was that the battery had no runner: twenty-two self-running files, two mutually incompatible summary dialects ("%d passed, %d skipped, %d failed" and "{n}/{m} passed"), and a total added up by hand. My own hand count had test_exact_closure.py at 20 when it is 26.

tools/run_battery.py counts the per-test PASS/FAIL/SKIP lines every file already emits, so it is dialect-independent, and it distinguishes a crashed file from a failing assertion:

crashed = proc.returncode != 0 and tally["FAIL"] == 0

The prior memo now carries an erratum and pins its numbers to its own slice instead of tracking the tree, since a memo that silently follows the code is how the drift arose.

5. Legacy Node harness quarantined

Moved to legacy/node-harness/ and excluded from the packet by SKIP_DIRS, enforced by G4. Three reasons, all documented in legacy/node-harness/README.md: the package name collided with the traaviis console script in pyproject.toml; npm test was red from a clean extraction because of a hard-coded absolute path to a sibling repository; and the description no longer matched what TRAAVIIS is.

test/fixtures/fake-claude.sh is needed by both batteries. It stays where the Python battery owns it and the quarantined test reaches up for it, so the dependency points legacy → live and never the reverse.

6. EvaluationV1.aggregation_profile

{"profile": "traaviis.aggregation.mean-over-scored.v1",
 "statistic": "arithmetic_mean", "population": "scored_episodes",
 "unscored_policy": "excluded", "scored": 3, "unscored": 1}

The report now states the denominator it used. Laws E14 and E15 pin that an unscored episode is excluded rather than imputed as zero — the difference is invisible in a mean but is the whole difference between "we did not score this" and "this scored nothing".

7. EvaluationV1.runtime_context

{"registry_version": "...", "wiring": "registry",
 "verifiers_available": [...], "verifier_versions": {...}}

Attests the registry that actually ran. wiring is "registry" or "caller_supplied", so a report cannot quietly present injected verifiers as the default set. Per the ruling, no free-form notes field: registry notes are deliberately not copied through.

8. Cross-extractor deterministic packet identity

Proved at the root and not merely through the gate. Extracting with zipfile without applying the manifest modes — so fake-claude.sh lands as -rw-r--r--, the exact case that used to break — and rebuilding reproduces the delivered hash. G7 then does it from both extractors on every acceptance run.

The first genuine repair task

residency-repair, a second template on the residency.repository.v1 substrate.

spec/contract.md   requires the module to return 2
src/mod.py         returns 1
target test        baseline exits 1  ->  patched exits 0
health control     baseline exits 0  ->  patched exits 0

This is the distinction the ruling was after. Every earlier Residency fixture answers "did the agent produce an admissible patch?". evidence-residency seeds return 1 and accepts either return 1 or return 2, so its acceptance test cannot fail on the subject it ships with — correctly reclassified as a conformance fixture. This one answers "did the agent fix the bug?", and the difference is visible in the artifacts rather than only in the prose.

Nine laws in test/test_repair_task.py, written mostly against the ways the shape can be faked:

R6 is the phase asymmetry on the case that motivates it: seed the subject already fixed while the plan still declares the target red, and the baseline contradicts the task. That is a statement about the fixture, so it must be error with reward: None, never fail.

R7 and R8 are the two ways a "repair" can be admissible and still not be one. nofix patches return 1 → return 3: it applies cleanly, cites the contract verbatim, and preserves the frozen world's sealed identity. Under evidence-residency that scores full marks. Here the tests cap holds it at 0.4. gutspec genuinely satisfies the target — it really does leave return 2 on disk — and is stopped only by the health control.

Two defects found on the way

The sealed environment is not a channel

The stub agent originally selected its mode from TRAAVIIS_REPAIR_MODE in the host environment. All three modes produced a byte-identical episode_id. That is runner._seal_env working as designed: the child environment is built solely from the task's agent_run_policy.environment and the host's is never inherited (§10a). A fixture cannot smuggle configuration through os.environ.

The mode now rides in the agent's argv, which is also the more honest model: these are three candidate agents answering one task, so the task bytes — and therefore task-… — must be identical across all three. R7 now asserts exactly that (bad["task_id"] == good["task_id"], distinct episode_ids), so the comparison is between two candidates and not two different tasks.

Multi-file unified diffs did not apply — a production defect

gutspec needs to patch two files. It could not, and the failure was not in the fixture.

traaviis/patchapply.py drove its hunk-body loop off each line's first character:

while i < n and lines[i] and lines[i][0] in " -+\\":

The next file's section header is --- a/spec/contract.md / +++ b/..., which by leading character alone are a removal and an addition. The applier swallowed the following header into the previous hunk and then reported a removed-line mismatch in the previous file. Multi-file diffs were systematically unrepresentable — while the module docstring already contemplated multiple sections ("A file may appear in at most one section").

The counts, not the leading character, terminate a hunk in unified diff. That is now what the applier does, which dissolves the ambiguity at the root. Because terminating on counts would let a body longer than its declared counts be silently skipped by the outer scan, a guard keeps that inadmissible:

if i < n and lines[i] and not lines[i].startswith("--- "):
    raise PatchError(...)

Four regression laws in test/test_patchapply.py, including the converse — a hunk that deletes a line whose text happens to look like a diff header must still work, or the fix would have traded one ambiguity for another.

This was reachable from the CLI by any agent emitting a two-file patch, so it is worth flagging as more than a test-support fix.

Decisions taken autonomously

Flagged for overturning.

1. The repair mode is an agent argument, not task configuration. The

alternative — putting it in agent_run_policy.environment — would move task-… per candidate and make R7's comparison meaningless. I judged one task with three candidate agents to be the correct model.

2. `patchapply` was fixed rather than the fixture reduced to one file. A

single-file gutspec cannot express "satisfy the target while wrecking the repository", which is the law R7 exists to state. The defect was real and in production code, so I fixed it and added regression laws.

3. Terminating a hunk on counts tightened admission slightly. Content between

a completed hunk and the next section is now rejected outright, where the outer scan previously skipped it. This would reject a diff --git preamble between sections. That seemed right for an applier documented as "evidence-grade — exact, fully-admitted, no fuzz", but it is a judgement call.

4. The prior memo's counts were pinned, not updated. Rewriting 340 to 358

would make the memo track the tree, which is the drift failure mode. It now carries a note pointing here.

5. `_ALL_PASS` is used from the test. R3 needs to ask "is this test red by the

default rule?", which means judging a command with its phase expectations stripped. It reaches for a private name to do so. Superseded: the reviewer suggested a public predicate instead. substrate_verifiers.command_set_passed(state) now exists and R3 uses it. See TRAAVIIS_WIRING_AND_COMPARE_CLOSURE_MEMO.md.

What was not built

Per the ruling, untouched: eval-…, agent-…, ORS, trvs serve --ors/--mcp, MCP, the REPL.

Next in the ruled order: trvs compare, then serial batch, then bundle-… distribution identity, then serve/ORS.