devlyn-cli 2.3.0 → 2.3.1

package/benchmark/auto-resolve/fixtures/F11-batch-import-all-or-nothing/NOTES.md

@@ -64,7 +64,12 @@ forces invariant derivation — the discriminating axis.
 
 ## Rotation trigger
 
-Retire when both arms consistently land > 90 across two shipped versions,
-OR when "all-or-nothing batch" becomes a recognized pattern such that
-solo arm reliably validates-first on the initial implementation pass.
+Headroom run `20260507-f10-f11-tier1-full-pipeline` rejected this fixture as
+pair-lift evidence: bare scored 98 and solo_claude scored 97. Keep it as an
+atomic batch control unless the visible contract is reworked to expose lower
+bare/solo ceilings.
+
+Retire when both `bare` and `solo_claude` consistently land > 90 across two
+shipped versions, OR when "all-or-nothing batch" becomes a recognized pattern
+such that solo arm reliably validates-first on the initial implementation pass.
 Whichever comes first.

package/benchmark/auto-resolve/fixtures/F12-webhook-raw-body-signature/NOTES.md

@@ -77,7 +77,13 @@ keywords. Raw-body trap is intentionally left without explicit
 
 ## Rotation trigger
 
-Retire when both arms consistently land > 90 across two shipped versions
-on this fixture. If the raw-body verifier (#5) becomes saturated faster
-than the others, replace it with a different platform blindspot rather
-than retiring the whole fixture.
+Headroom run `20260511-f12-webhook-headroom` rejected this fixture as pair-lift
+evidence: bare scored 85 and solo_claude scored 99. Bare still missed one of
+seven verifiers, but the `bare` and `solo_claude` judge scores exceed the
+headroom ceilings. Keep it as a webhook/security control unless the visible
+contract is reworked to expose lower bare/solo ceilings.
+
+Retire when both `bare` and `solo_claude` consistently land > 90 across two
+shipped versions on this fixture. If the raw-body verifier (#5) becomes
+saturated faster than the others, replace it with a different platform blindspot
+rather than retiring the whole fixture.

package/benchmark/auto-resolve/fixtures/F15-frozen-diff-race-review/NOTES.md

@@ -92,7 +92,13 @@ that doesn't read the await sequence carefully will gloss over these.
 
 ## Rotation trigger
 
-Retire when both bare and solo arms consistently land > 85 across two
-shipped versions. If 2026 baseline reliably catches the awaited RMW race
-on cold read of someone else's code, the frozen-diff review thesis also
-needs updating — not just the seeded bug.
+Headroom run `20260511-f15-concurrency-headroom` rejected this fixture as
+pair-lift evidence: bare scored 99 and solo_claude scored 94, so `bare` and
+`solo_claude` are above the headroom ceilings (`bare <= 60`, `solo_claude <=
+80`). Keep the fixture as a frozen-diff review control unless the visible
+contract is reworked to expose a lower solo ceiling.
+
+Retire when both `bare` and `solo_claude` consistently land > 85 across two
+shipped versions. If 2026 baseline reliably catches the awaited RMW race on
+cold read of someone else's code, the frozen-diff review thesis also needs
+updating — not just the seeded bug.

package/benchmark/auto-resolve/fixtures/F16-cli-quote-tax-rules/NOTES.md

@@ -19,6 +19,18 @@ F1/F2 test CLI shape, but not business-rule arithmetic. F10/F11/F12 test
 server behavior and persistence. F15 tests review behavior. None combine
 hidden product math, exact machine output, and source-of-truth pricing.
 
+## Measurement status
+
+Pair evidence passed in `20260510-f16-f23-f25-combined-proof`:
+bare `50`, solo_claude `75`, pair `96`, margin `+21`, wall `1.28x`,
+arm `l2_risk_probes`, verdict `pair_evidence_passed`.
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss duplicate-SKU aggregation
+before stock validation and exact integer tax/shipping totals; observable
+command `node "$BENCH_FIXTURE_DIR/verifiers/exact-success.js"` exposes the miss.
+
 ## Retirement
 
 Retire or replace this fixture if both bare and solo consistently score

package/benchmark/auto-resolve/fixtures/F16-cli-quote-tax-rules/spec.md

@@ -55,3 +55,9 @@ the output must be machine-readable.
 - A quote over combined stock exits `2`, prints one JSON error to stderr, and prints no stdout.
 - The stock error object includes `sku`, `available`, and `requested`.
 - `git diff --stat` shows only `bin/cli.js` and `tests/cli.test.js` touched (the pricing seed comes from setup, not the arm).
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss duplicate-SKU aggregation
+before stock validation or tax/discount calculation; observable command
+`node "$BENCH_FIXTURE_DIR/verifiers/exact-success.js"` exposes the miss.

package/benchmark/auto-resolve/fixtures/F2-cli-medium-subcommand/NOTES.md

@@ -39,16 +39,19 @@ small enough that every arm can plausibly finish in < 10 minutes.
 
 ## When should this fixture be retired or replaced?
 
-When both arms score > 95 for two consecutive shipped versions — i.e., the
-fixture saturates and no longer differentiates. Candidate replacement: a
-similar-size CLI task with multiple interacting flags or a subcommand that
-spawns a child process.
+When both `bare` and `solo_claude` score > 95 for two consecutive shipped
+versions — i.e., the fixture saturates and no longer differentiates. Candidate
+replacement: a similar-size CLI task with multiple interacting flags or a
+subcommand that spawns a child process.
 
 ## Calibration history
 
 - v3.4   skill 57 / bare 45 / margin +12 (gpt-5.3-codex judge)
 - v3.4.1 skill 59 / bare 43 / margin +16 (gpt-5.3-codex judge)
 - v3.5   skill 92 / bare 81 / margin +11 (gpt-5.4 xhigh judge) — huge absolute jump; bare silent-catch caught
+- 20260512-f2-medium-headroom bare 83 / solo_claude 95 — rejected as
+  pair-lift evidence because both baseline scores exceed current headroom
+  ceilings.
 
 Absolute scores jumped with the stronger judge. Margin stays solid (+11
 after stdlib calibration is expected to open a few points more).

package/benchmark/auto-resolve/fixtures/F21-cli-scheduler-priority/NOTES.md

@@ -20,6 +20,18 @@ F16 covers checkout arithmetic. F10/F11/F12/F15 cover server behavior. None
 exercise a CLI algorithm where the correct result depends on sorting,
 interval arithmetic, and output ordering at once.
 
+## Measurement status
+
+Pair evidence passed in `20260511-f21-current-riskprobes-v1`: bare `33`,
+solo_claude `66`, pair `99`, margin `+33`, wall `1.47x`,
+arm `l2_risk_probes`, verdict `pair_evidence_passed`.
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss global priority ordering
+combined with blocked-interval earliest-fit placement; observable command
+`node "$BENCH_FIXTURE_DIR/verifiers/priority-blocked.js"` exposes the miss.
+
 ## Retirement
 
 Retire or replace when both bare and solo consistently exceed the headroom

package/benchmark/auto-resolve/fixtures/F21-cli-scheduler-priority/spec.md

@@ -59,3 +59,9 @@ failure reasons must be deterministic.
 - Unknown resources are reported in `rejected` without aborting the whole run.
 - Duplicate request ids are invalid input: exit `2`, one JSON error to stderr, no stdout.
 - `git diff --stat` shows only `bin/cli.js` and `tests/cli.test.js` touched.
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss global priority ordering
+across resources while preserving blocked-interval earliest-fit placement;
+observable command `node "$BENCH_FIXTURE_DIR/verifiers/priority-blocked.js"` exposes the miss.

package/benchmark/auto-resolve/fixtures/F22-cli-ledger-close/NOTES.md

@@ -20,6 +20,14 @@ F16 covers order quote arithmetic, but not ledger idempotency or full-input
 validation before mutation. F21 covers interval scheduling. Server fixtures
 cover API behavior rather than CLI reconciliation.
 
+## Measurement status
+
+Headroom runs reject F22 as full-pipeline pair-lift evidence. In
+`20260507-f21-f22-full-pipeline-pair`, F22 scored bare 91 / solo_claude 98 and
+failed the headroom gate. In `20260508-f22-exact-error-headroom`, F22 scored
+bare 94 / solo_claude 98 after the exact-error fixture revision. Keep it as a ledger
+reconciliation control, not as counted `solo < pair` evidence.
+
 ## Retirement
 
 Retire or replace if both bare and solo consistently score above the headroom

package/benchmark/auto-resolve/fixtures/F23-cli-fulfillment-wave/NOTES.md

@@ -20,6 +20,18 @@ F21 covers interval scheduling. F16 covers quote arithmetic. F22 was too easy
 for bare in the first calibration run. This fixture targets allocation rollback
 and inventory consumption across multiple dimensions.
 
+## Measurement status
+
+Pair evidence passed in `20260510-f16-f23-f25-combined-proof`: bare `33`,
+solo_claude `66`, pair `97`, margin `+31`, wall `2.25x`,
+arm `l2_risk_probes`, verdict `pair_evidence_passed`.
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss all-or-nothing rollback after
+a higher-priority order consumes stock first; observable command
+`node "$BENCH_FIXTURE_DIR/verifiers/priority-rollback.js"` exposes the miss.
+
 ## Retirement
 
 Retire or replace if both bare and solo consistently exceed the headroom

package/benchmark/auto-resolve/fixtures/F23-cli-fulfillment-wave/spec.md

@@ -68,3 +68,9 @@ orders must be deterministic.
 - Lot choice is FEFO by expiry date, then lot id.
 - `remaining` is sorted by warehouse id, then sku, then expires, then lot.
 - `git diff --stat` shows only `bin/cli.js` and `tests/cli.test.js` touched.
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss all-or-nothing rollback after
+a higher-priority order tentatively allocates FEFO lots; observable command
+`node "$BENCH_FIXTURE_DIR/verifiers/priority-rollback.js"` exposes the miss.

package/benchmark/auto-resolve/fixtures/F25-cli-cart-promotion-rules/NOTES.md

@@ -17,12 +17,24 @@ examples rather than only checking a happy path.
 ## Why existing fixtures do not cover it
 
 F16 covers quote tax rules, but not multiple line-promotion types plus an order
-coupon. F21/F23 cover scheduling/allocation but became oracle-control fixtures.
+coupon. F21/F23 cover scheduling/allocation, not checkout interaction ordering.
 This fixture keeps the F16-style fair visible-contract shape while testing a
 different checkout interaction.
 
+## Measurement status
+
+Pair evidence passed in `20260510-f16-f23-f25-combined-proof`: bare `25`,
+solo_claude `75`, pair `99`, margin `+24`, wall `1.65x`,
+arm `l2_risk_probes`, verdict `pair_evidence_passed`.
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss the interaction between
+duplicate aggregation, line promotions, tax base, coupon order, and shipping;
+observable command `node "$BENCH_FIXTURE_DIR/verifiers/exact-success.js"` exposes the miss.
+
 ## Retirement
 
-Retire or replace this fixture if bare or solo consistently reaches ceiling, or
-if a later fixture covers the same promotion-order and catalog-source failure
-mode with cleaner full-pipeline lift.
+Retire or replace this fixture if either `bare` or `solo_claude` consistently
+reaches ceiling, or if a later fixture covers the same promotion-order and
+catalog-source failure mode with cleaner full-pipeline lift.

package/benchmark/auto-resolve/fixtures/F25-cli-cart-promotion-rules/spec.md

@@ -62,3 +62,10 @@ and stdout must stay machine-readable.
 - The stock error object includes `sku`, `available`, and `requested`.
 - Changing `data/catalog.json` prices or rates changes command output without code changes.
 - `git diff --stat` shows only `bin/cli.js` and `tests/cli.test.js` touched (the catalog seed comes from setup, not the arm).
+
+## Solo-headroom hypothesis
+
+A capable solo_claude baseline is expected to miss the interaction between
+duplicate-SKU aggregation, line-promotion ordering, coupon ordering, and
+tax/shipping thresholds; observable command
+`node "$BENCH_FIXTURE_DIR/verifiers/exact-success.js"` exposes the miss.

package/benchmark/auto-resolve/fixtures/F26-cli-payout-ledger-rules/NOTES.md

@@ -15,11 +15,17 @@ adversarial ledger examples with repeated IDs, refunds, disputes, and reserves.
 ## Why existing fixtures do not cover it
 
 F16 covers quote math and F25 covers cart promotions, but neither has ledger
-idempotency or conflicting duplicate events. F21/F23 became oracle-control
-fixtures, so this adds a fresh visible-contract stateful arithmetic candidate.
+idempotency or conflicting duplicate events. F21/F23 cover scheduling and
+allocation ordering, not payout ledger arithmetic.
+
+## Measurement status
+
+Headroom run `20260508-f26-headroom` rejected F26 as full-pipeline pair-lift
+evidence: bare scored 25, but `solo_claude` scored 98 and passed all 4
+verification commands, so the fixture is at solo ceiling. Keep it as a ledger
+math control unless the spec is revised to expose a lower solo ceiling.
 
 ## Retirement
 
-Retire or replace this fixture if solo consistently reaches ceiling or if
-another fixture provides the same idempotent-ledger signal with cleaner
-full-pipeline pair lift.
+Retire or replace this fixture if another fixture provides the same
+idempotent-ledger signal with cleaner full-pipeline pair headroom.

package/benchmark/auto-resolve/fixtures/F3-backend-contract-risk/NOTES.md

@@ -25,4 +25,11 @@ update tests but forget backward-compat requirements (single-item route,
 
 ## Rotation trigger
 
-Retire when both arms consistently score > 95 AND produce 2+ new tests covering paging edge cases without pipeline intervention.
+Headroom run `20260511-f3-http-error-headroom` rejected this fixture as
+pair-lift evidence after the invalid-query HTTP error body verifier was
+tightened: bare scored 97 and solo_claude scored 99. Keep it as a backend
+contract control unless the visible contract is reworked to expose lower
+bare/solo ceilings.
+
+Retire when both `bare` and `solo_claude` consistently score > 95 AND produce
+2+ new tests covering paging edge cases without pipeline intervention.

package/benchmark/auto-resolve/fixtures/F3-backend-contract-risk/expected.json

@@ -26,10 +26,12 @@
       "stdout_not_contains": []
     },
     {
-      "cmd": "node -e 'const { app } = require(\"./server\"); const http = require(\"http\"); const s = http.createServer(app).listen(0, () => { const { port } = s.address(); http.get(`http://127.0.0.1:${port}/items?per_page=abc`, r => { console.log(r.statusCode); s.close(); }); });'",
+      "cmd": "node -e 'const { app } = require(\"./server\"); const http = require(\"http\"); const s = http.createServer(app).listen(0, () => { const { port } = s.address(); http.get(`http://127.0.0.1:${port}/items?per_page=abc`, r => { let b = \"\"; r.on(\"data\", c => b += c); r.on(\"end\", () => { const d = JSON.parse(b); console.log(JSON.stringify({ status: r.statusCode, error: d.error, field: d.field })); s.close(); }); }); });'",
       "exit_code": 0,
       "stdout_contains": [
-        "400"
+        "\"status\":400",
+        "\"error\":\"invalid_query\"",
+        "\"field\":\"per_page\""
       ],
       "stdout_not_contains": []
     }

package/benchmark/auto-resolve/fixtures/F3-backend-contract-risk/spec.md

@@ -50,6 +50,6 @@ so existing assertions continue to pass alongside new paging assertions.
 - Server start: `node server/index.js` listens on port 3000 (exit via SIGINT).
 - `curl -s http://127.0.0.1:3000/items | jq '.total'` returns `2`.
 - `curl -s 'http://127.0.0.1:3000/items?per_page=1&page=2' | jq '.items[0].name'` returns `"beta"`.
-- `curl -s 'http://127.0.0.1:3000/items?per_page=abc' -o /dev/null -w '%{http_code}'` returns `400`.
+- `curl -s 'http://127.0.0.1:3000/items?per_page=abc'` returns HTTP status `400` with JSON error body `{ "error": "invalid_query", "field": "per_page" }`.
 - `node --test tests/server.test.js` passes; must include ≥ 2 new paging tests.
 - `git diff --stat` shows only `server/index.js` and `tests/server.test.js` touched.

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/NOTES.md

@@ -0,0 +1,34 @@
+# F31 CLI seat rebalance
+
+## Failure mode
+
+This fixture detects implementations that pass simple entitlement updates while
+missing the interaction between priority ordering, transfer rollback, rejected
+row ordering, and exact machine-readable error handling.
+
+## Pipeline phase target
+
+PLAN must preserve the ordering distinction between processing order and
+rejected-output order. IMPLEMENT must keep transfer mutations all-or-nothing.
+VERIFY should build adversarial cases where a later high-priority transfer
+changes the outcome of an earlier low-priority reserve, and where a failed
+transfer would corrupt state if mutations are applied too early.
+
+## Why existing fixtures do not cover it
+
+F21 covers scheduling priority and blocked intervals. F23 covers inventory
+allocation rollback. F25 covers checkout calculation order. This fixture covers
+account entitlement reconciliation with a different state shape and a duplicate
+event-id hard error.
+
+## Retirement
+
+Headroom run `20260512-f31-seat-rebalance-headroom` rejected this fixture as
+pair-lift evidence: bare scored 33 but carried judge/result/verify
+disqualifiers, and solo_claude scored 98 with all 3 verification commands
+passing. It should remain a control fixture unless reworked to lower the solo
+ceiling.
+
+Retire or replace this fixture if either `bare` or `solo_claude` consistently
+reaches ceiling, or if a later fixture covers priority event processing plus
+all-or-nothing transfer rollback with cleaner full-pipeline lift.

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/expected.json

@@ -0,0 +1,57 @@
+{
+  "verification_commands": [
+    {
+      "cmd": "node --test tests/cli.test.js",
+      "exit_code": 0,
+      "stdout_contains": [],
+      "stdout_not_contains": ["not ok "]
+    },
+    {
+      "cmd": "node \"$BENCH_FIXTURE_DIR/verifiers/priority-transfer-rollback.js\"",
+      "exit_code": 0,
+      "stdout_contains": ["\"ok\":true"],
+      "stdout_not_contains": [],
+      "contract_refs": [
+        "Process events globally by `priority` descending, then `effective_at` ascending, then `id` ascending.",
+        "`transfer` is all-or-nothing. It accepts only when both accounts exist, the source has `used >= qty`, the destination has at least `qty` free seats, and both accounts have the same `region` unless `allow_cross_region` is `true`.",
+        "A rejected `transfer` must not change either account. Use reason `unknown_account`, `region_mismatch`, `insufficient_used`, or `no_capacity` for the first failing transfer rule in the order listed above.",
+        "`rejected` is ordered in the original input event order. Each row has keys `id`, `reason`.",
+        "`accounts` is sorted by account id ascending. Each row has keys `id`, `region`, `seats`, `used`, `free`, where `free = seats - used`.",
+        "On success, write exactly one JSON object to stdout and no stderr. Keys: `applied`, `rejected`, `accounts`."
+      ]
+    },
+    {
+      "cmd": "node \"$BENCH_FIXTURE_DIR/verifiers/duplicate-event-error.js\"",
+      "exit_code": 0,
+      "stdout_contains": ["\"ok\":true"],
+      "stdout_not_contains": [],
+      "contract_refs": [
+        "Before processing any event, duplicate event ids are invalid input: exit `2`, write exactly one JSON error object `{ \"error\": \"duplicate_event_id\", \"id\": string }` to stderr, and write no stdout."
+      ]
+    }
+  ],
+  "forbidden_patterns": [
+    {
+      "pattern": "catch\\s*\\([^)]*\\)\\s*\\{[^}]*return\\s+(null|undefined|''|\\{\\})",
+      "description": "silent catch returning fallback in seat rebalance path",
+      "files": ["bin/cli.js", "tests/cli.test.js"],
+      "severity": "disqualifier"
+    },
+    {
+      "pattern": "catch\\s*\\([^)]*\\)\\s*\\{\\s*\\}",
+      "description": "empty catch block",
+      "files": ["bin/cli.js", "tests/cli.test.js"],
+      "severity": "disqualifier"
+    },
+    {
+      "pattern": "/\\*\\s*eslint-disable",
+      "description": "eslint-disable without scoped justification",
+      "files": ["bin/cli.js"],
+      "severity": "disqualifier"
+    }
+  ],
+  "required_files": ["bin/cli.js", "tests/cli.test.js"],
+  "forbidden_files": [],
+  "max_deps_added": 0,
+  "spec_output_files": ["bin/cli.js", "tests/cli.test.js"]
+}

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/metadata.json

@@ -0,0 +1,10 @@
+{
+  "id": "F31-cli-seat-rebalance",
+  "category": "high-risk",
+  "difficulty": "high",
+  "timeout_seconds": 1500,
+  "required_tools": ["node"],
+  "browser": false,
+  "deps_change_expected": false,
+  "intent": "Add a bench-cli rebalance-seats command that reads account capacity and seat events from a JSON file, processes events by priority with all-or-nothing transfers, rejects invalid per-event operations without corrupting state, and prints exact applied, rejected, and final account rows."
+}

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/setup.sh

@@ -0,0 +1,2 @@
+#!/usr/bin/env bash
+set -e

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/spec.md

@@ -0,0 +1,67 @@
+---
+id: "F31-cli-seat-rebalance"
+title: "Seat rebalance command"
+status: planned
+complexity: high
+depends-on: []
+---
+
+# F31 Seat rebalance command
+
+## Context
+
+`bench-cli` currently has greeting and version commands only. The task:
+add a `rebalance-seats` command that reads account capacity and seat events from
+a JSON file, processes events by priority with all-or-nothing transfers, rejects
+invalid per-event operations without corrupting state, and prints exact applied,
+rejected, and final account rows.
+
+This is account entitlement reconciliation. Downstream billing tools parse the
+output, so success and error output must be exact machine-readable JSON.
+
+## Requirements
+
+- [ ] `bench-cli rebalance-seats --input <path>` reads JSON shaped as `{ "accounts": [{ "id": string, "region": string, "seats": number, "used": number }], "events": [event] }`.
+- [ ] Valid event types are `reserve`, `release`, and `transfer`.
+- [ ] `reserve` events have keys `id`, `type`, `account`, `qty`, `priority`, and `effective_at`.
+- [ ] `release` events have keys `id`, `type`, `account`, `qty`, `priority`, and `effective_at`.
+- [ ] `transfer` events have keys `id`, `type`, `from`, `to`, `qty`, `priority`, `effective_at`, and optional `allow_cross_region`.
+- [ ] Before processing any event, duplicate event ids are invalid input: exit `2`, write exactly one JSON error object `{ "error": "duplicate_event_id", "id": string }` to stderr, and write no stdout.
+- [ ] Before processing any event, account rows must have unique ids, non-empty string `id` and `region`, integer `seats >= 0`, and integer `used` with `0 <= used <= seats`. Invalid account input exits `2` with one JSON error object and no stdout.
+- [ ] Before processing any event, every event `qty` must be a positive integer, every `priority` must be an integer, and every `effective_at` must be a non-empty string. Invalid event input exits `2` with one JSON error object and no stdout.
+- [ ] Process events globally by `priority` descending, then `effective_at` ascending, then `id` ascending.
+- [ ] `reserve` accepts only when the account exists and has at least `qty` free seats. On accept, increase that account's `used` by `qty`. Otherwise reject the event with reason `unknown_account` or `no_capacity`.
+- [ ] `release` accepts only when the account exists and `used >= qty`. On accept, decrease that account's `used` by `qty`. Otherwise reject the event with reason `unknown_account` or `insufficient_used`.
+- [ ] `transfer` is all-or-nothing. It accepts only when both accounts exist, the source has `used >= qty`, the destination has at least `qty` free seats, and both accounts have the same `region` unless `allow_cross_region` is `true`.
+- [ ] A rejected `transfer` must not change either account. Use reason `unknown_account`, `region_mismatch`, `insufficient_used`, or `no_capacity` for the first failing transfer rule in the order listed above.
+- [ ] On success, write exactly one JSON object to stdout and no stderr. Keys: `applied`, `rejected`, `accounts`.
+- [ ] `applied` is ordered in processing order. Each row has keys `id`, `type`.
+- [ ] `rejected` is ordered in the original input event order. Each row has keys `id`, `reason`.
+- [ ] `accounts` is sorted by account id ascending. Each row has keys `id`, `region`, `seats`, `used`, `free`, where `free = seats - used`.
+- [ ] `tests/cli.test.js` is updated. Existing tests still pass and at least two new tests cover `rebalance-seats`: one successful priority/transfer scenario and one validation failure.
+
+## Constraints
+
+- **No new npm dependencies.**
+- **No hidden mutable global state.** The command must derive output only from the input JSON for that invocation.
+- **No silent catches.** Parse and file-read failures must emit a visible JSON error to stderr and exit `2`.
+- **No extra stdout/stderr text** on the success path; downstream tooling parses stdout as JSON.
+
+## Out of Scope
+
+- Persisting account state between command invocations.
+- Adding billing invoices, plan catalogs, or currency calculations.
+- Adding web UI or server routes.
+- Touching `server/`, `web/`, or `tests/server.test.js`.
+
+## Verification
+
+- `node --test tests/cli.test.js` exits 0.
+- A later high-priority transfer is processed before an earlier low-priority reserve, and the low-priority reserve can lose capacity because of that ordering.
+- A rejected transfer leaves both source and destination account usage unchanged.
+- Region mismatch rejects a transfer unless `allow_cross_region` is `true`.
+- `rejected` rows are reported in the original input event order, even though processing order is priority based.
+- Duplicate event ids exit `2`, print exactly `{ "error": "duplicate_event_id", "id": string }` to stderr, and print no stdout.
+- Final `accounts` rows are sorted by id and include exact `free` values.
+- `git diff --stat` shows only `bin/cli.js` and `tests/cli.test.js` touched.
+- Solo-headroom hypothesis: solo_claude is expected to miss transfer rollback or rejected-row ordering under priority processing; observable command `node "$BENCH_FIXTURE_DIR/verifiers/priority-transfer-rollback.js"` exposes the miss.

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/task.txt

@@ -0,0 +1,7 @@
+Add a bench-cli rebalance-seats command that reads account capacity and seat events from a JSON file, processes events by priority with all-or-nothing transfers, rejects invalid per-event operations without corrupting state, and prints exact applied, rejected, and final account rows.
+
+The command should be `bench-cli rebalance-seats --input <path>`. The input JSON has account rows with id, region, seats, and used, plus event rows for reserve, release, and transfer operations. Process all events by priority descending, then effective_at ascending, then id ascending.
+
+Transfers must be all-or-nothing: if either account is missing, the source lacks used seats, the destination lacks free seats, or the regions differ without allow_cross_region, reject the transfer and leave both accounts unchanged. Reserve and release should reject per-event failures without aborting the whole command. Duplicate event ids are invalid input and must exit 2 with exactly one JSON error object on stderr and no stdout.
+
+On success, stdout must be exactly one JSON object with applied rows in processing order, rejected rows in original input order, and final account rows sorted by account id with free = seats - used. Update `tests/cli.test.js` so existing tests still pass and at least two new tests cover rebalance-seats, including one successful priority/transfer scenario and one validation failure. Do not add dependencies or touch server/web files.

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/verifiers/duplicate-event-error.js

@@ -0,0 +1,35 @@
+'use strict';
+const assert = require('node:assert');
+const { spawnSync } = require('node:child_process');
+const fs = require('node:fs');
+const os = require('node:os');
+const path = require('node:path');
+
+const work = process.env.BENCH_WORKDIR || process.cwd();
+const cli = path.join(work, 'bin', 'cli.js');
+const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'f31-duplicate-'));
+const input = path.join(tmp, 'events.json');
+
+fs.writeFileSync(input, JSON.stringify({
+  accounts: [
+    { id: 'team-a', region: 'us', seats: 5, used: 1 }
+  ],
+  events: [
+    { id: 'dup', type: 'reserve', account: 'team-a', qty: 1, priority: 2, effective_at: '2026-01-01T09:00:00Z' },
+    { id: 'dup', type: 'release', account: 'team-a', qty: 1, priority: 1, effective_at: '2026-01-01T09:01:00Z' }
+  ]
+}), 'utf8');
+
+const result = spawnSync('node', [cli, 'rebalance-seats', '--input', input], {
+  cwd: work,
+  encoding: 'utf8'
+});
+
+assert.strictEqual(result.status, 2);
+assert.strictEqual(result.stdout, '');
+assert.deepStrictEqual(JSON.parse(result.stderr), {
+  error: 'duplicate_event_id',
+  id: 'dup'
+});
+
+console.log(JSON.stringify({ ok: true }));

package/benchmark/auto-resolve/fixtures/F31-cli-seat-rebalance/verifiers/priority-transfer-rollback.js

@@ -0,0 +1,53 @@
+'use strict';
+const assert = require('node:assert');
+const { spawnSync } = require('node:child_process');
+const fs = require('node:fs');
+const os = require('node:os');
+const path = require('node:path');
+
+const work = process.env.BENCH_WORKDIR || process.cwd();
+const cli = path.join(work, 'bin', 'cli.js');
+const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'f31-rebalance-'));
+const input = path.join(tmp, 'events.json');
+
+fs.writeFileSync(input, JSON.stringify({
+  accounts: [
+    { id: 'team-a', region: 'us', seats: 5, used: 3 },
+    { id: 'team-b', region: 'us', seats: 4, used: 1 },
+    { id: 'team-eu', region: 'eu', seats: 4, used: 0 }
+  ],
+  events: [
+    { id: 'low-reserve', type: 'reserve', account: 'team-b', qty: 3, priority: 1, effective_at: '2026-01-01T09:00:00Z' },
+    { id: 'bad-cross', type: 'transfer', from: 'team-a', to: 'team-eu', qty: 1, priority: 8, effective_at: '2026-01-01T09:02:00Z' },
+    { id: 'high-transfer', type: 'transfer', from: 'team-a', to: 'team-b', qty: 2, priority: 10, effective_at: '2026-01-01T09:05:00Z' },
+    { id: 'after-release', type: 'release', account: 'team-a', qty: 1, priority: 7, effective_at: '2026-01-01T09:03:00Z' },
+    { id: 'after-reserve', type: 'reserve', account: 'team-a', qty: 5, priority: 6, effective_at: '2026-01-01T09:04:00Z' }
+  ]
+}), 'utf8');
+
+const result = spawnSync('node', [cli, 'rebalance-seats', '--input', input], {
+  cwd: work,
+  encoding: 'utf8'
+});
+assert.strictEqual(result.status, 0, result.stderr || result.stdout);
+assert.strictEqual(result.stderr, '');
+const parsed = JSON.parse(result.stdout);
+
+assert.deepStrictEqual(parsed, {
+  applied: [
+    { id: 'high-transfer', type: 'transfer' },
+    { id: 'after-release', type: 'release' },
+    { id: 'after-reserve', type: 'reserve' }
+  ],
+  rejected: [
+    { id: 'low-reserve', reason: 'no_capacity' },
+    { id: 'bad-cross', reason: 'region_mismatch' }
+  ],
+  accounts: [
+    { id: 'team-a', region: 'us', seats: 5, used: 5, free: 0 },
+    { id: 'team-b', region: 'us', seats: 4, used: 3, free: 1 },
+    { id: 'team-eu', region: 'eu', seats: 4, used: 0, free: 4 }
+  ]
+});
+
+console.log(JSON.stringify({ ok: true }));

package/benchmark/auto-resolve/fixtures/F32-cli-subscription-renewal/NOTES.md

@@ -0,0 +1,38 @@
+# F32 Subscription renewal command
+
+## Failure mode
+
+This fixture targets billing-style state mutation where an implementation can
+look correct on isolated cases but fail the interaction between renewal
+priority, tentative credit application, rollback after `payment_required`, exact
+credit consumption order, and strict JSON row shapes.
+
+## Pipeline phase coverage
+
+- PLAN must preserve the exact input/output field names and ordering clauses.
+- RISK_PROBES should derive a compound priority + rollback + shape probe.
+- IMPLEMENT must avoid input-order processing and must not leak tentative credit
+  consumption from rejected renewals.
+- VERIFY pair mode should catch aliased keys, extra keys, and weak tests that
+  check only one field rather than the full parsed output.
+
+## Why existing fixtures do not cover it
+
+F25 covers pricing math and output shape, F31 covers entitlement transfers and
+duplicate-id errors, and F23 covers fulfillment rollback. F32 combines billing
+credits with a failed high-priority renewal that must roll back before a later
+renewal can consume credits, plus exact nested output key sets.
+
+## Retirement criteria
+
+Retire or replace this fixture if both `bare` and `solo_claude` score above 95
+on two current-model runs, or if another active fixture covers priority-ordered
+tentative monetary credit rollback with exact nested output shape and duplicate
+ID error contracts.
+
+## Pair-candidate status
+
+Rejected as pair-lift evidence by `20260512-f32-subscription-renewal-headroom`:
+bare scored 33, but solo_claude scored 98 and passed all 3 verification
+commands. Keep it as a billing rollback/shape control, not as a pair arm target,
+unless it is reworked and clears a fresh headroom gate.