@zerwiz/ymir 0.1.16 → 0.1.18

package/assets/reference/docs/verification/dispatch-auth.md

@@ -0,0 +1,204 @@
+# Dispatch authentication verification
+
+Audience: maintainer verification.
+
+This record supports the dispatch judgment rules in `.agents/skills/quota-array-dispatch/SKILL.md` and the bounded vendor probe in `bin/fm-vendor-auth-probe.sh`.
+It records only facts that must be re-established when a producer or vendor version changes.
+Task chronology, incident transcripts, and credential metadata stay in private reports or PR evidence.
+
+Firstmate resolves a candidate's provider family, credential surface, and applicable quota by reading the evidence below and reasoning in the open.
+No script maps a model to a provider, a provider to a credential store, or a name prefix to a family, so the facts here are what that reasoning rests on.
+Credential paths below are shown with the home directory replaced by `<home>`.
+
+## Quota granularity the judgment depends on
+
+Verified 2026-07-30 against quota-axi 0.1.16 for the provider and model-scope relationships below.
+That release's captured default output included `quotaSemantics.description`; the current default TOON and JSON fallback field placement are verified against 0.1.29 in the next section.
+Current dispatch reads the TOON scope and `limitedBy` fields; the JSON fallback's corresponding `scope` and `boundedBy` fields preserve the same provider/model applicability without relying on the `--full`-only description.
+
+```json
+{
+  "provider": "codex",
+  "state": { "status": "fresh", "stale": false },
+  "quotaSemantics": {
+    "status": "known",
+    "description": "Codex base account windows bound every model. Named model windows add bounds for that model; code-review windows describe a separate workload and are not included in model availability.",
+    "effectiveAvailability": [
+      { "scope": "all_models", "status": "known", "effectivePercentRemaining": 64, "boundedBy": ["weekly"] },
+      { "scope": "model:codex_bengalfox", "status": "known", "effectivePercentRemaining": 64, "boundedBy": ["weekly", "model:codex_bengalfox:7d"] }
+    ]
+  }
+}
+```
+
+Three properties follow and are load-bearing for dispatch:
+
+- An `all_models` (or `all_products`) scope is real evidence for every model in that provider family, including a model with no window of its own.
+- A `model:`-scoped entry is an additional bound for that one model. `model:codex_bengalfox` is the GPT-5.3-Codex-Spark window and bounds nothing else.
+- A named-model window can be tighter than the account bound, so it must not be read across models. In the same snapshot Claude reported `all_models` with `effectivePercentRemaining` 10 while `model:fable` reported 4, limited by the `model:fable` window itself. A non-Fable Claude model reads 10, not 4.
+
+`quotaSemantics.status` is `unknown` with no `effectiveAvailability` entries at all for providers whose vendor exposes no window (observed for `cursor` and `copilot`).
+`state.authStatus` is present only for some providers (observed for `grok` alone), so its absence is missing evidence, not a credential fault.
+
+## Completion-runway and selection shape the judgment depends on
+
+Verified 2026-08-18 against quota-axi 0.1.29 schema 5, captured from an isolated `quota-axi@0.1.29` install.
+The default TOON exposed these table headers, with row counts normalized to `N`:
+
+```text
+quota[N]{provider,scope,effectivePercentRemaining,spendPriority,runway,confidence,limitedBy,resetsAt}:
+exhaustion[N]{provider,scope,usableRunwaySeconds,projectedExhaustedAt,limitingWindowId}:
+attention[N]{provider,scope,kind,detail,remedy}:
+```
+
+`exhaustion[]` and `attention[]` are sparse, so an empty table is rendered with count zero and no row fields.
+The command below records the JSON fallback shape without persisting account-specific quota values:
+
+```sh
+quota-axi --json | jq '{schemaVersion, effectiveAvailabilityFields: ([.providers[]?.quotaSemantics.effectiveAvailability[]? | keys] | unique), runwayFields: ([.providers[]?.quotaSemantics.effectiveAvailability[]?.runway? | select(type == "object") | keys] | unique), selectionFields: ([.providers[]?.quotaSemantics.effectiveAvailability[]?.selection? | select(type == "object") | keys] | unique), paceFields: ([.providers[]?.quotaSemantics.effectiveAvailability[]?.pace? | select(type == "object") | keys] | unique), windowPaceFields: ([.providers[]?.windows[]?.pace? | select(type == "object") | keys] | unique)}'
+```
+
+```json
+{
+  "schemaVersion": 5,
+  "effectiveAvailabilityFields": [
+    [
+      "boundedBy",
+      "effectivePercentRemaining",
+      "limitingWindowIds",
+      "pace",
+      "runway",
+      "scope",
+      "selection",
+      "status"
+    ]
+  ],
+  "runwayFields": [
+    [
+      "projectionConfidence",
+      "status"
+    ]
+  ],
+  "selectionFields": [
+    [
+      "spendPriority",
+      "status"
+    ]
+  ],
+  "paceFields": [
+    [
+      "status",
+      "worstReservePercentPoints",
+      "worstReserveWindowId"
+    ]
+  ],
+  "windowPaceFields": [
+    [
+      "burnMultiple",
+      "reservePercentPoints",
+      "status"
+    ]
+  ]
+}
+```
+
+This live snapshot was all `through_reset`, so finite-runway fields were omitted.
+`usableRunwaySeconds`, `projectedExhaustedAt`, and `limitingWindowId` remain in default `--json` when `runway.status` is `projected_exhaustion` or `exhausted_now`.
+`selection.unmeasurableWindowIds`, scope `aheadWindowIds`/`unknownWindowIds`, and window `pace.reason` likewise remain in default `--json` when they apply.
+`quotaSemantics.description`, `behindWindowIds`, `onPaceWindowIds`, and per-window cycle-progress internals are `--full` only.
+There is no `projectionBasis` field; its absence means `cycle_average`.
+`runway` and `selection` are nested under each effective-availability scope, so the same provider/model applicability rules govern headroom, runway, and `spendPriority`.
+Projection confidence is not present on every known runway, so selection must preserve that absence as uncertainty rather than fabricate it.
+The older-schema fallback contract is owned by `quota-array-dispatch`; this evidence does not reinterpret an absent runway, pace, or selection field.
+
+## Provider-family counterfactual that this producer schema supports
+
+Verified 2026-07-30 on Pi 0.82.0 and quota-axi 0.1.16.
+
+```sh
+pi --list-models terra
+```
+
+```text
+provider      model          context  max-out  thinking  images
+openai-codex  gpt-5.6-terra  272K     128K     yes       yes
+```
+
+The Pi catalog is authoritative for Pi model support and reports the provider family in its own column.
+For `harness=pi`, `model=openai-codex/gpt-5.6-terra` the catalog establishes the model is supported and belongs to the `openai-codex` family, and the Codex `all_models` scope above supplies fresh, known 64 effective remaining for every model in that family.
+No Terra-specific window exists in the snapshot, and `quota-axi auth --json` lists no `pi:openai-codex` source.
+Both absences are missing model-level and source-level detail, not contradictory evidence, so this candidate is dispatchable with the model-level uncertainty disclosed.
+
+```sh
+pi --list-models gpt-9.9-nonexistent
+```
+
+```text
+No models matching "gpt-9.9-nonexistent"
+```
+
+A listing that reaches the account and returns no row is the authoritative negative that does block a candidate.
+
+## Credential sources are independent per provider
+
+Verified 2026-07-30 against quota-axi 0.1.16.
+
+`quota-axi auth --json` reports each provider's credential sources separately, which is what lets a candidate be scoped to the one surface it actually authenticates through:
+
+```json
+[
+  { "provider": "claude", "sources": [
+      { "source": "oauth-file", "path": "<home>/.claude/.credentials.json", "status": "missing" },
+      { "source": "keychain", "status": "available" } ] },
+  { "provider": "codex", "sources": [
+      { "source": "auth-json", "path": "<home>/.codex/auth.json", "status": "available" },
+      { "source": "cli-rpc", "path": "<path-to>/codex", "status": "available" } ] },
+  { "provider": "grok", "sources": [
+      { "source": "auth-json", "path": "<home>/.grok/auth.json", "status": "available" },
+      { "source": "pi:xai", "status": "available" } ] },
+  { "provider": "kimi", "sources": [
+      { "source": "pi:kimi-coding", "status": "available" },
+      { "source": "kimi-code-cli", "status": "expired", "error": "kimi_code_cli_credential_expired" } ] }
+]
+```
+
+Observed source statuses are `available`, `expired` (with an `error` slug), and `missing`.
+
+- A provider can carry a healthy source beside a missing or expired one, so a provider must not be collapsed to a single status. Claude's `oauth-file` is missing while its keychain source is available, and Kimi's standalone CLI credential is expired while its Pi source is available.
+- A `pi:`-prefixed source exists only where Pi holds its own credential for that family (`pi:xai`, `pi:kimi-coding`). Pi's `openai-codex` family has none, because it authenticates through the Codex store that the `codex` provider already lists. A missing `pi:` source is therefore never evidence against a Pi candidate.
+
+Neither this per-source shape nor `state.authStatus` exists before quota-axi 0.1.16.
+`bin/fm-bootstrap.sh` enforces the current compatibility floor through `bin/fm-quota-axi-lib.sh`.
+
+Grok also reports `credits.remaining: 0` alongside `percentRemaining: 41` on a healthy account.
+That zero is a prepaid balance, not the subscription window, and is never headroom.
+
+## Standalone Grok discovery probe
+
+Verified 2026-07-30 on `grok 0.2.117 (f1c06093089f) [stable]`.
+
+```sh
+grok --version
+grok models   # stdin closed, single attempt, hard-bounded
+```
+
+Observed:
+
+- `grok models` exits `0` and its first stdout line is `You are logged in with grok.com.` for an authenticated session.
+- With a home directory holding no Grok credential, the first stdout line is `You are not authenticated.`, also with exit status `0`.
+- Because the status is `0` in both cases, the exit status is not a verdict; only the literal first stdout line is examined, and a blank first line does not authenticate.
+- `<home>/.grok/auth.json` was byte-identical across the authenticated run (`mtime`, `size`, and mode `0600` unchanged), so the probe is a read in that path.
+
+These discriminator strings are un-owned vendor UI text.
+`bin/fm-vendor-auth-probe.sh` pins the verified version, reports `versionVerified=no` when the running CLI differs, and classifies any unrecognized first line as `indeterminate` rather than authenticated.
+Re-run the two commands above and update this section and the pinned version together when the vendor CLI changes.
+
+## Regression coverage
+
+`tests/fm-vendor-auth-probe.test.sh` drives the real script against a fake vendor CLI that records every invocation's argv and anything readable on stdin.
+It asserts that the script accepts no harness, model, or provider input, never calls `quota-axi`, exits alike for every probe result because it renders no verdict, invokes only the two fixed non-destructive argv forms with stdin closed, holds a real bound even when the configured bound is zero or malformed, and never echoes raw vendor output.
+`tests/fm-spawn-dispatch-profile.test.sh` owns spawn's deterministic profile and harness refusals.
+`tests/fm-bootstrap.test.sh` owns the quota-axi version-floor diagnostic.
+`tests/fm-quota-array-dispatch-live-e2e.test.sh` drives the public Pi skill-loading interface against one fake schema-5 snapshot per case, served as quota-axi's default TOON.
+It covers TOON-first `spendPriority` ranking among candidates that pass eligibility, reasoning-class, and runway-feasibility gates, explicit accounting for unmeasurable runway, the strongest-reasoning constraint, and the runway feasibility floor over a higher `spendPriority`.
+The skill's primary path is that default TOON; `--json` is the documented defensive fallback, and this section records the producer `--json` shape that fallback consumes.

package/assets/reference/docs/verification/muse.md

@@ -0,0 +1,221 @@
+# Verification: the muse (Muse Code) crewmate adapter
+
+Active empirical evidence for firstmate's muse adapter.
+The skill tree rooted at [`.agents/skills/harness-adapters/SKILL.md`](../../.agents/skills/harness-adapters/SKILL.md) owns the operating facts; this record owns how they were established and what is still unproven.
+
+## Subject
+
+| Field | Value |
+|---|---|
+| Version | `Muse Code 0.1.0 (0.1.0-R708.1)`, build sha `427a430436` |
+| Verified | 2026-08-05, extended 2026-08-06 with the credentialed multi-step smoke |
+| Artifact | `muse-aarch64-macos`, sha256 `4290bfafa5bbb81a6fd493aaea12f848c789b1d22edfa0c4b849151deba3e70c` |
+| Platform | macOS arm64 (Darwin 25.5.0) |
+
+The binary was fetched from the published channel and its checksum matched the published manifest before any run:
+
+```
+$ curl -sS 'https://api.meta.ai/muse-code/channels/muse-stable'
+{"channel":"muse-stable","version":"0.1.0-R708.1",...,"state":"public","min_version":null}
+
+$ shasum -a 256 muse-bin
+4290bfafa5bbb81a6fd493aaea12f848c789b1d22edfa0c4b849151deba3e70c  muse-bin
+```
+
+Every run below used an isolated `XDG_CONFIG_HOME` and `XDG_DATA_HOME` in a scratch directory and a throwaway git workspace, driven through tmux the way firstmate drives a crewmate pane.
+`install.sh` was deliberately bypassed, so no shell profile and no `~/.local/bin` entry on the host was touched.
+
+## What the model provider limits
+
+Live TUI and session behavior below was observed against the built-in `--provider echo` startup provider, except for the provider-authentication prompt.
+The credential paths and unauthenticated wait were probed separately against the default `meta` provider.
+Turn-boundary structure, the trust dialog, interrupt, exit, composer rendering, credential behavior, and the event-log schema are real and verified.
+Busy-state behavior under a genuine multi-step, real-model tool loop was verified separately on 2026-08-06 against the default `meta` provider with a live model, and is recorded under [the credentialed multi-step smoke](#the-credentialed-multi-step-smoke-verified-2026-08-06).
+
+## Verified facts
+
+### Process identity
+
+The published launcher `exec`s a version-suffixed binary, so the live process name changes on every auto-update:
+
+```
+$ grep -nE 'muse-bin|exec ' launcher.sh
+969:  candidate="$work/muse-bin"
+977:  target="$dir/muse-bin-$version"
+1035:  printf '%s/muse-bin-%s\n' "$dir" "$version"
+1135:  exec "$binary" "$@"
+```
+
+`ps -o comm= -p <pid>` returns the full executable path, whose basename is `muse-bin-<version>`.
+That is why both `bin/fm-harness.sh` and `bin/backends/tmux.sh` match the anchored prefix `muse-bin-*` rather than an exact name, and why neither can rely on an install-path component: `~/.local/bin/muse-bin-<version>` contains no `muse` path component.
+The Muse launch clears `CLAUDECODE`, `PI_CODING_AGENT`, `GROK_AGENT`, `FM_PI_HARNESS`, `CURSOR_AGENT`, and `CURSOR_INVOKED_AS` before the worker starts so foreign primary markers cannot override the versioned ancestry.
+
+[`runtime-backends.md`](runtime-backends.md#agent-liveness-name-sources) owns the resulting tmux liveness verdict and its relationship to the portable decoy regression.
+
+### Turn lifecycle
+
+A two-turn session produced exactly two run brackets, the second closed by an Escape interrupt:
+
+```
+9  {"kind":"run","run_id":"d352a097-...","event":{"kind":"started","prompt":"hello from firstmate"}}
+45 {"kind":"run","run_id":"d352a097-...","event":{"kind":"terminal","terminal":"completed","turn_duration_ms":8152}}
+49 {"kind":"run","run_id":"b50dac92-...","event":{"kind":"started","prompt":"second turn to interrupt"}}
+78 {"kind":"run","run_id":"b50dac92-...","event":{"kind":"terminal","terminal":"cancelled","reason":"cancelled during model step"}}
+```
+
+The log's first record carries the workspace binding key:
+
+```
+"payload_type": "runtime.session.metadata",
+"payload": {"kind":"metadata","record":{"workspace_root":".../muselab/ws1","provider_id":"echo",...}}
+```
+
+The fold transitions live, sampled during a 25-second in-flight turn:
+
+```
+T+ 5s fold=busy
+T+10s fold=busy
+T+15s fold=busy
+T+20s fold=busy
+T+25s fold=busy
+T+30s fold=settled
+```
+
+Two decoys were observed in real logs and are pinned by regressions in `tests/fm-muse-harness.test.sh`:
+a nested `"record":{"kind":"terminal"}` cleanup-effect payload that is not a run terminal, and independent sub-agent run lifecycles under `subagent/<child-session-id>/session.jsonl`.
+The same regression suite verifies that unique resolution is cached, a changed current-day main-session namespace restores ambiguity to unknown, a replacement spawn binding selects its fresh main log, missing cached logs fail closed, and cached sub-agent paths are rejected.
+
+### Autonomy, trust, and sandbox
+
+A fresh untrusted workspace shows the trust dialog with option 1 preselected:
+
+```
+Do you trust this workspace?
+> 1  Trust and continue
+  2  Quit
+Use Up/Down or 1/2, then Enter. Esc quits.
+```
+
+`--yolo` suppresses it entirely and the status bar reports `echo · <workspace> · YOLO`.
+This matters because approval and the sandbox are ON by default and `--sandbox-network` defaults to `proxy-only`, which the binary reports as requiring managed shell sandboxing - a crewmate needs ordinary git and network access.
+
+### Credentials
+
+`muse auth set --provider` accepts only `meta`.
+An unauthenticated launch does not exit; it waits indefinitely:
+
+```
+  Sign in at this page:
+    https://auth.meta.com/oauth/device/?code=DGXZ-NRPR
+  Waiting for approval…
+  Esc cancel
+```
+
+That is why `bin/fm-spawn.sh` preflights worker-reachable `META_API_KEY` or `<config>/muse/auth.json` and refuses before creating an endpoint.
+A caller-only `META_API_KEY` is refused because a long-lived backend daemon does not inherit it, while the non-secret `XDG_CONFIG_HOME` and `XDG_DATA_HOME` roots are resolved to absolute paths before preflight and forwarding so the stored credential and session-log binding reach the same worker environment.
+
+### Foreign personal context
+
+The interactive TUI rejects the `exec`-only flag:
+
+```
+$ muse --no-foreign-personal-context --provider echo hi
+invalid TUI options: error: unexpected argument '--no-foreign-personal-context' found
+  tip: a similar argument exists: '--no-session-log'
+```
+
+`MUSE_EXPERIMENTAL_FOREIGN_PERSONAL_CONTEXT_KILL` is the control that works in TUI mode.
+Comparing the `context_block_diagnostic` block ids emitted by otherwise identical runs, with the operator's real `~/.claude` rules present and no project `AGENTS.md`:
+
+```
+base     blocks=rules_file,workspace_identity,security_mode,skills_catalog,session_identity,subagent_delegation
+killon   blocks=workspace_identity,security_mode,skills_catalog,session_identity,subagent_delegation
+kill1    blocks=workspace_identity,security_mode,skills_catalog,session_identity,subagent_delegation
+```
+
+Repeating the comparison with a project `AGENTS.md` present confirms the kill switch drops only the FOREIGN rules:
+
+```
+a4base   blocks=rules_file,workspace_identity,security_mode,skills_catalog,session_identity,subagent_delegation
+a4kill   blocks=rules_file,workspace_identity,security_mode,skills_catalog,session_identity,subagent_delegation
+```
+
+The `tui.foreign_context_notice_shown` flag in `settings.json` suppresses only the notice, never the loading, so a quiet later launch is not evidence of a clean context.
+
+### Composer rendering
+
+Captured with `tmux capture-pane -p -e`:
+
+```
+^[[38;2;90;160;255m^[[48;2;38;56;84m⟩ ^[[38;2;204;211;219mhello from firstmate^[[39m
+^[[0m^[[38;2;90;160;255m⟩ ^[[39m
+```
+
+Prompt glyph `⟩` (U+27E9) at luminance ~149.9 against the 128 default ghost threshold; typed text at ~209.8.
+After a single Escape the interrupted prompt is restored into the composer at the same bright ~209.8, and `C-u` clears it.
+
+## The credentialed multi-step smoke (verified 2026-08-06)
+
+This was the one item deferred until a `META_API_KEY` was available, because it is what decides whether a settled log may classify `idle`.
+An open run was always positive proof of a turn in flight, but a settled log only proves no run is open at that instant, so the classifier held idle behind an opt-in in case a real turn spanned several runs.
+The smoke below answered that: one run brackets a whole multi-step turn, and an Escape interrupt closes that run with `terminal=cancelled` rather than leaving the turn to continue in another run.
+The credentialed result gives a settled Muse log the same idle trust as the Claude and Pi push sources, so the opt-in was removed and `bin/fm-busy-lib.sh` classifies a settled log `idle` outright.
+Muse auto-updates its vendor binary underneath the fleet, firstmate normalizes the versioned process identity to the `muse` harness before busy classification, and the session log's own metadata carries semver `0.1.0` plus a build SHA that cannot be matched to that normalized identity.
+A verified-build allowlist against this coarse identity would be false precision because it could not distinguish the running build, as well as a maintenance treadmill against the auto-updating binary.
+
+Both runs below used the default `meta` provider with model `muse-spark-1.2-contributor`, on a real firstmate-launched crewmate pane, authenticated through the stored `~/.config/muse/auth.json` written by `muse auth set --provider meta --api-key-stdin` so the key never entered `argv`.
+
+### One turn stays inside one run
+
+A single 8-step tool loop (shell, file reads, a file write, a shell append) ran as one submitted turn in session `629b3bc1-5dd7-4a0d-a901-69701850922c`, log `~/.local/share/muse/sessions/2026/08/06/629b3bc1-5dd7-4a0d-a901-69701850922c/session.jsonl`.
+The whole 828-record turn is bracketed by exactly one run pair, 23 tool batches deep:
+
+```
+$ grep -cE '"kind":"run","run_id":"[^"]*","event":\{"kind":"started"' session.jsonl
+1
+$ grep -cE '"kind":"run","run_id":"[^"]*","event":\{"kind":"terminal"' session.jsonl
+1
+$ grep -c '"payload_type":"tool_batch.effect.started"' session.jsonl
+23
+
+10  {"kind":"run","run_id":"db5869ed-...","event":{"kind":"started","prompt":"...launch-brief..."
+827 {"kind":"run","run_id":"db5869ed-...","event":{"kind":"terminal","terminal":"completed",
+      "reason":null,"turn_duration_ms":75243,"time_to_first_token_ms":69583,"eot_gate_ms":3907}
+```
+
+Scope the count to `"kind":"run"` as above.
+A bare `grep -c '"event":{"kind":"started"'` returns 56 on the same log, because every tool batch effect reuses that inner event shape.
+
+### Busy sampling and interrupt
+
+Session `e4e0b4f4-38d0-46dc-b669-dfb5de92e0e0` sampled the fold while a multi-step turn was in flight, then interrupted it with Escape mid tool loop.
+Five consecutive samples of `fm_busy_muse_run_state` on the bound log returned `busy`, and `fm_busy_classify` returned `busy muse-session-log` for the same samples; the fold settled immediately after the interrupt.
+Its run closed as cancelled rather than staying open:
+
+```
+10  {"kind":"run","run_id":"a098d532-...","event":{"kind":"started","prompt":"...launch-brief..."
+103 {"kind":"run","run_id":"a098d532-...","event":{"kind":"terminal","terminal":"cancelled",
+      "reason":"cancelled during model step","turn_duration_ms":7849}
+```
+
+That is the same terminal shape the `echo`-provider interrupt produced, now confirmed against a live model mid tool loop.
+
+`tests/fm-muse-harness.test.sh` pins the resulting classifier behavior: a log settled by either terminal reads `idle`, an open run reads `busy`, and only a resolution failure reads `unknown`.
+
+## Refreshing this record
+
+Run both opt-in live guards after any muse upgrade, because the version-suffixed process name, session protocol, and styled composer are vendor-controlled surfaces:
+
+```
+FM_HARNESS_LIVENESS_DRIFT=1 bin/fm-test-run.sh tests/fm-harness-liveness-drift-live-e2e.test.sh
+FM_MUSE_SIGNALS_LIVE=1 bin/fm-test-run.sh tests/fm-muse-signals-live-e2e.test.sh
+```
+
+The Muse signals guard requires a real `muse` binary and tmux but uses `--provider echo`, so it does not require `META_API_KEY` and cannot re-check the real-model turn-to-run relationship on its own.
+The guard follows SGR state through the final prompt glyph and rejects both bright-then-dark and malformed-RGB negative controls before accepting that glyph's effective luminance.
+
+muse's launcher can replace the running binary underneath the fleet, so an upgrade that changes the session protocol also invalidates the credentialed evidence above.
+Repeat that smoke after a protocol-affecting upgrade: run one real multi-step tool-loop turn with credentials in place, confirm the run-scoped `started`/`terminal` counts are still exactly one each, and confirm an Escape still yields `terminal` with `cancelled`.
+A build that ever split one turn across several runs would make a settled log ambiguous, which is a classifier change rather than a note in this file.
+
+The portable counterparts that run in ordinary CI are `tests/fm-muse-harness.test.sh`, `tests/fm-tmux-agent-liveness.test.sh`, `tests/fm-composer-lib.test.sh`, and `tests/fm-composer-ghost.test.sh`.