docs: expand practical test coverage matrix

docs/implementation-plan.v1.md

@@ -189,7 +189,7 @@ developers call rather than duplicating orchestration in workflow YAML:
 scripts/test-unit                 [--package PATTERN] [--run REGEXP] [--race]
 scripts/test-integration          [--suite NAME|all] [--run-id ID] [--resume]
 scripts/test-e2e                  [--scenario NAME|all] [--run-id ID] [--resume]
-scripts/test-env                  doctor|status|logs|collect|recover|reuse|stop|reset|purge|gc ...
+scripts/test-env                  doctor|coverage|status|logs|collect|recover|reuse|stop|reset|purge|gc ...
 scripts/windows/test-host.ps1     Prepare|Status|Run|Collect|Stop|Reset
 ```
 
@@ -286,6 +286,8 @@ and preserves only the manifest plus state required for `status`, `logs`, or
 
 - `doctor [--e2e]` is read-only and validates local/container prerequisites and,
   when requested, the native Windows runner before allocating a run;
+- `coverage` is read-only and validates the coverage inventory against listed
+  unit/integration/E2E cases and their required host/layer metadata;
 - `status RUN_ID` is read-only and reports steps, health, disk usage, owned
   Compose/Windows resources, and whether resume is valid;
 - `logs RUN_ID [COMPONENT]` reads bounded/tail output; `collect RUN_ID` writes a
@@ -344,6 +346,425 @@ stable nonzero exit codes, and a final summary giving the run ID, seed, report
 path, resume command when applicable, and exact cleanup command. Document the
 same workflow in `docs/testing.md` during Phase 0.
 
+### 2.5 Traceable practical-scenario coverage catalogue
+
+Create `test/coverage.toml` as the executable coverage inventory. Every concrete
+case has a stable ID, class (`happy`, `hostile-recovery`, or `race`), owning
+requirement/plan section, layer, suite, supported platforms, privilege/session
+fixture, speed class, implementation phase, and case name. The ranges below are
+shorthand; the inventory expands them into one row per case. `scripts/test-env
+coverage` rejects duplicate IDs, missing case implementations, cases that cannot
+be listed by their suite, and normative requirements with no test. Test code
+references the ID in its name or metadata, and failures print it.
+
+Use line coverage as a warning signal, not as a substitute for scenarios.
+Nevertheless, pure policy/domain/protocol/config packages must maintain at least
+90% statement coverage and core store/session/runtime packages at least 80%.
+Platform syscall wrappers report coverage separately and are judged primarily by
+native integration cases. Any excluded unreachable/generated/error branch needs
+a reviewed explanation in the inventory. CI reports coverage changes by package;
+a new critical branch may not lower coverage merely because the repository-wide
+percentage remains above threshold.
+
+For every numeric/time/size limit, generate `minimum`, `minimum+1`, normal,
+`maximum-1`, `maximum`, and `maximum+1` cases where meaningful, plus zero,
+negative decoded values where the source type permits them, integer overflow,
+and unit-conversion boundaries. For every state machine, exhaustively generate
+all state/event pairs and then add the concurrent interleavings listed below.
+Pairwise generation covers independent dimensions such as shell, source type,
+foreground/background, privilege intent, online/offline state, and output kind;
+use a full Cartesian product only where interactions affect an invariant.
+
+#### Happy-path cases
+
+- `HP-CFG-01..12` — Parse every annotated server/client/Windows TOML; apply
+  compiled defaults and explicit flag precedence; normalize Unicode/space-
+  containing absolute paths; validate Unix and Windows shell paths; advertise
+  only current-platform shells; render redacted effective configuration; run
+  `--check-config` without opening listeners/stores; create the first Windows
+  template; start service live-but-not-ready for placeholder routing; restart
+  into ready state after valid configuration; preserve config across upgrade.
+- `HP-PROTO-01..10` — Negotiate exact and overlapping protocol ranges; ignore
+  unknown optional fields; round-trip every envelope and control message;
+  transfer maximum legal binary output/script chunks; Zstandard round-trip
+  empty/compressible/incompressible data; accept Unicode command/output bytes;
+  preserve client-observed and server-receipt timestamps; encode/decode every
+  structured error and optional field without losing presence; store/forward the
+  single `elevated` request bit unchanged while accepting Windows attempted/
+  effective contexts only as client-originated result metadata.
+- `HP-CTL-01..18` — List/get clients and commands; run foreground/background
+  command text and scripts; print UUID before following; resume follow; paginate
+  commands and byte slices; select stdout/stderr; append text/raw/file stdin;
+  close stdin; send TERM/KILL successfully to Windows and parse every portable
+  signal name for later Unix delivery; inspect Windows identity,
+  expiry, truncation, incomplete output, incidents, and takeover; repair and
+  acknowledge incidents; exercise matching gRPC and JSON-RPC unary behavior;
+  accept optional global `--request-id` and ignore it for reads.
+- `HP-IDEM-01..10` — Generate UUIDv7 at the correct owner; accept externally
+  supplied UUIDv7; return the same result for identical run/stdin/close/signal/
+  incident/takeover retries; map run request ID directly to `issue_uuid`; retain
+  one mutation result through transport reconnect; suppress a matching command
+  tombstone replay; preserve UUIDv7 FIFO order for equal-millisecond issuance.
+- `HP-SES-01..14` — Register a new client; reconnect the same durable instance;
+  fence its old connection; accept a different instance when none is live;
+  authorize and consume an exact live-instance takeover; expire an unused grant;
+  advertise/reconcile capacity; dispatch into running and queued slots; requeue
+  a transient rejection; resume event and script delivery; reconcile active and
+  terminal-unacknowledged records; discard a server-confirmed terminal client
+  record; keep multiple clients fair.
+- `HP-CLIENT-01..14` — Create and retain one durable client-instance ID; start
+  with an empty spool; durably admit and acknowledge commands; advertise queue/
+  running capacity; dequeue fairly; assign and replay event sequences; compact
+  acknowledged events; retain terminal-unacknowledged records; discard them only
+  after reconciliation/acknowledgement; enforce local command/aggregate quotas;
+  restart with clean recovery; reconnect with saved backoff/session inputs; stop
+  all supervised Jobs on orderly service shutdown.
+- `HP-CMD-01..18` — Execute `cmd` and PowerShell command text and uploaded
+  scripts; use foreground/background modes; apply requested/omitted CWD;
+  materialize identity-scoped default CWD; overlay an empty/small/large legal
+  environment; handle empty/multiline text, shell metacharacters, PowerShell
+  Unicode, and space/non-ASCII executable/CWD/wrapper paths; return exit codes 0
+  and nonzero; record every public lifecycle; run 1 and configured maximum
+  concurrent commands; queue then start work; execute short and descendant-
+  producing commands; collect resource snapshots; apply every supported
+  individual/composed resource profile before release.
+- `HP-WINCTX-01..14` — With one active standard user run normal as
+  `ACTIVE_USER`; with a split-token administrator run normal filtered and
+  elevated through the linked full token; accept an already-full admin token;
+  build/verify a restricted medium token when UAC is off; run elevated as
+  `ACTIVE_SYSTEM` after linked-token unavailability; use `LOCAL_SYSTEM` as final
+  elevated fallback; with no active session run normal as `LOCAL_SERVICE` and
+  elevated as `LOCAL_SYSTEM`; persist attempts/effective SIDs/session; omit
+  session fields for Session 0; load/unload active profile; build the effective
+  environment; access ACL-permitted drive, localhost, and network resources.
+- `HP-LAUNCH-01..16` — Authenticate the per-command launcher pipe; create the
+  launcher and shell suspended; assign the Job before release; persist/flush
+  prepared then authorized state; resume exactly once; capture both streams;
+  write and close stdin; keep descendants in the Job; observe root and complete-
+  tree exit; drain pipes to EOF; deliver CTRL_BREAK through the verified helper;
+  escalate TERM after grace; terminate immediately on KILL; enforce Job limits;
+  remove wrappers/scripts after terminal acknowledgement; preserve effective
+  identity on later lifecycle events.
+- `HP-OUT-01..14` — Preserve binary/invalid-UTF-8 bytes, empty writes, partial
+  lines, no-newline output, long lines, interleaved stdout/stderr, exact chunk
+  boundaries, cumulative event acknowledgements, reconnect replay, byte-offset
+  pagination within an event, live follow after historical output, normal drain,
+  explicit incomplete marker, and normal client/server compression reuse.
+- `HP-STORE-01..20` — Initialize/migrate SQLite; append inline and segmented
+  payloads; group commit; reopen cleanly; query stable cursors; charge command/
+  client/server bytes; rotate output window; evict whole oldest terminal UUID;
+  protect active commands; reserve closeout; reclaim terminal age; create and
+  cap tombstones; rotate audit by bytes/age; maintain incident history; run safe
+  repair; acknowledge loss; derive/clear dirty state; checkpoint WAL; backup and
+  restore a quiesced store.
+- `HP-FLOW-01..12` — Send heartbeat while idle; reset activity on any valid
+  frame/Pong; reset reconnect backoff after stability; keep Ping/Pong/Close and
+  acknowledgements ahead of saturated data; apply fair writer/persistence
+  scheduling; cross high then low watermarks; retain assigned events; summarize
+  unsequenced client loss; use closeout reserve; enforce per-command/session send
+  windows; reconnect with deterministic full jitter.
+- `HP-SVC-01..16` — Idempotently install/start/stop/restart/uninstall the SCM
+  service; boot before login; preserve data on uninstall; change Automatic/
+  Manual in SCM; start one tray per session from the Run key; reconnect tray
+  after Explorer restart; exit tray without stopping service; read status/log;
+  perform an elevated configuration action; rotate service logs; operate without
+  tray; run Session 0 on Server Core; attach console for human CLI modes; display
+  native error/dialog when no parent console exists.
+- `HP-OPS-01..12` — Expose liveness during recovery and readiness afterward;
+  keep healthy scopes usable; emit bounded-cardinality metrics; rotate/redact
+  logs and audits; report storage and cache usage; perform orderly shutdown;
+  resume after service/server restart; validate deployment examples; upgrade
+  with config check/snapshot/migration; restore prior version/config on an
+  intentionally failed pre-migration check; generate checksums/package metadata.
+- `HP-HARNESS-01..14` — Doctor a cold and cached environment; create/list a run;
+  execute one case/scenario; collect JSON/JUnit/report; stop and resume; recover
+  an interrupted controller; reuse immutable definitions into fresh state;
+  reset; purge one run; dry-run and execute age GC; purge all confirmed RVBox
+  runs; retain caches by default; explicitly clear caches/images; preserve file
+  ownership and unrelated labelled fixtures.
+
+#### Bad, hostile, and abnormal-recovery cases
+
+- `BH-CFG-01..24` — Reject missing/unknown/duplicate keys, wrong TOML types,
+  invalid UTF-8, bare/overflow/negative durations, illegal zeros, limit and
+  watermark inversions, incompatible profiles, relative paths, missing/wrong-
+  type shell files, changed shell identity before launch, disallowed CWD,
+  insecure state/config/work ownership or ACL, symlink/reparse-point escape,
+  reserved device/ADS/NT-object path misuse, disallowed or inaccessible UNC
+  paths, bad URL/TLS name, listener collision,
+  malformed Windows quoting, and non-loopback JSON-RPC without explicit enable;
+  prove no listener, store mutation, or child starts after static failure.
+- `BH-PROTO-01..28` — Reject text WebSocket messages, invalid fragmentation,
+  empty/unknown payload, incompatible major/minor, absent/wrong fencing fields,
+  oversize envelope/spec/control/HTTP body/chunk/script, protobuf recursion/
+  length abuse, integer overflow, invalid enum/oneof combinations, malformed
+  UUID/client ID, invalid declared compressed sizes, truncated/corrupt Zstandard,
+  compression bombs, impossible event sequence/revision, changed duplicate
+  payload, out-of-order/overlap/hole script chunks, post-commit chunks, SHA/size
+  mismatch, invalid stdin sequence, and unsupported signal; close/fence only the
+  offending peer and keep daemons alive.
+- `BH-TLSNET-01..18` — Reject wrong CA/name/expired/not-yet-valid certificate,
+  plaintext endpoint, wrong nginx path/upgrade, and unreachable proxy; recover
+  from DNS refusal, TCP reset, TLS failure, proxy restart, half-open connection,
+  packet loss/duplication/reordering, extreme latency, slow reads/writes,
+  permanently blocked writer, write deadline, missing Pong, clock jump, and
+  reconnect storm without goroutine/file-descriptor growth or data loss.
+- `BH-CTL-01..24` — Reject malformed/empty/batch JSON-RPC where unsupported,
+  unknown method, invalid params/base64/timestamp, oversized body, tampered or
+  filter-mismatched cursor, page-size overflow, unknown client/command/incident,
+  illegal lifecycle action, stdin after close/terminal, signal before eligible
+  state, takeover for wrong/stale instance, repair of irreparable evidence,
+  acknowledgement without/beyond note bound, reused request ID with changed
+  method/target/content, read-only request ID leakage into domain, and client
+  disconnect during follow; never cancel work merely because the caller exits.
+- `BH-SES-01..24` — Reject stale/unknown generation, stale output after fencing,
+  live different-instance collision without exact grant, mismatched/expired/
+  consumed grant, capacity lies/overflow, dispatch to wrong generation,
+  permanent client rejection, reconnect snapshot missing accepted/running state,
+  matching tombstone against stale server state, hash/revision/terminal
+  contradiction, incomplete reconciliation snapshot, client dirty-store claim,
+  lost/repeated result, offline target, queue capacity exhaustion, and one noisy/
+  corrupt client; isolate the scope and keep unrelated clients/control responsive.
+- `BH-CLIENT-01..28` — Handle missing/corrupt/permission-denied client-instance
+  identity, duplicate daemon lock, spool schema/checksum/segment corruption,
+  uncommitted spool tail, missing acknowledged bytes, stale accepted record,
+  incomplete prepared/authorized process metadata, unknown live process, local
+  queue full, per-command/aggregate spool full, client disk floor, script temp
+  orphan, tombstone overflow, event-sequence counter drift, impossible server Ack,
+  dirty local storage, server rejecting registration/version, repeated connect
+  refusal, shutdown deadline, service kill, and config/shell change across
+  restart. Do not claim a complete reconciliation snapshot or accept work until
+  essential local state is trustworthy or explicitly resolved.
+- `BH-IDEM-01..16` — Reject noncanonical/non-v7 IDs, all-zero/wrong-length binary
+  forms, same UUID with changed immutable execution spec/script bytes, same
+  request ID across mutation kinds/targets, duplicate sequence with changed
+  bytes, and replay older than full history but still in tombstones; define and
+  test best-effort behavior after tombstone horizon expiry; never create a second
+  process or mutation result on any retained duplicate.
+- `BH-SCRIPT-01..18` — Treat filename as display only; reject path separators,
+  traversal, reserved names, NUL/control abuse, excessive metadata, changed
+  replay bytes, sparse/overlapping writes, premature commit, digest mismatch,
+  quota exhaustion, disk-full, permission/share violation, wrapper replacement,
+  symlink/reparse race, shell executable replacement, and cancellation during
+  upload; clean only owned temporary files and preserve durable rejection detail.
+- `BH-WINCTX-01..26` — Handle no linked token, standard user elevation request,
+  Administrator Protection/approval-only policy, disabled/removed account,
+  locked/disconnected/logging-off session, `WTSQueryUserToken` failure, multiple
+  ambiguous active sessions, console-session change, session-ID reuse with a new
+  logon SID, token SID/session/integrity/type mismatch, restricted-token creation
+  failure, profile load/unload failure, environment block failure, LocalService
+  logon failure, `TokenSessionId` failure, active-SYSTEM failure, final LocalSystem
+  failure, explicit CWD denial, identity-work-root tampering, absent mapped drive/
+  HKCU/network credentials, and Session 0 interactive API failure. Follow only
+  the allowed pre-launch chain and emit a bounded decision record with no
+  effective context if exhausted.
+- `BH-LAUNCH-01..34` — Reject spoofed/remote/early launcher or signal-helper pipe
+  clients, wrong PID/creation time/SID/session/generation/checksum/length,
+  inherited unrelated handles, launcher outside Job, breakaway attempt, shell/
+  wrapper/CWD/profile validation failure, invalid or case-colliding Windows
+  environment keys, oversized environment block, unencodable wrapper content,
+  process creation/assignment/resume failure, helper attach/delivery refusal,
+  PID reuse, root disappearance, child
+  holding pipes forever, descendant escape attempt, output read/write error,
+  stdin broken pipe, Job accounting failure, unsupported Job limit, service death
+  before/after preparation/authorization, launcher death at every handshake, and
+  incomplete EOF. Before authorization reject/cancel safely; after durable
+  authorization terminate/interrupt and never redispatch.
+- `BH-OUTFLOW-01..24` — Survive output flood from one/many commands, tiny writes,
+  incompressible data, invalid bytes, compressor error, client raw backlog,
+  offline spool full, pinned send window full, server ingress full, command/
+  client/global quota full, physical disk floor, loss-marker reservation
+  pressure, missing EOF, corrupt spool record, repeated Ack, Ack beyond sent
+  sequence, server retention removing resume point, and slow follower. Preserve
+  lifecycle/control progress and expose every loss/incomplete range honestly.
+- `BH-STORE-01..36` — Recover uncommitted tails and marked eviction; detect short
+  committed files, checksum/frame corruption, missing segment, wrong file type,
+  path replacement, SQLite corruption/lock/busy/readonly, WAL/shm loss, schema or
+  migration checksum mismatch, partial migration, fsync/rename/directory-sync
+  failure, ENOSPC at every append/closeout/audit phase, counter/reservation drift,
+  orphan temp/tombstone files, interrupted audit/age rotation, tombstone cap,
+  filesystem free-space breach, permission change, backup inconsistency, and
+  restore mismatch. Gate only affected scopes, expose incidents promptly, and
+  require acknowledgement before clearing known loss.
+- `BH-RET-01..20` — Reject new command when its own cap cannot fit; reject one
+  client/global admission when active/non-evictable data alone fills the tier;
+  evict terminal commands with no output; charge scripts/stdin/metadata/history/
+  idempotency data; never partly evict a command; preserve active commands,
+  tombstones, audit, unresolved incidents, assigned client events, and closeout
+  reserve; handle no eligible victim; retain truncation metadata; apply 30-day
+  and independent audit age/byte rotation exactly once.
+- `BH-SVC-01..28` — Reject non-admin install/config/uninstall, untrusted binary/
+  config path, direct internal service/launcher/helper mode, wrong SCM launch,
+  duplicate/corrupt service registration, insecure ProgramData ACL, malicious
+  tray pipe client, spoofed admin claim, cross-session tray action, oversized/
+  stalled IPC, second tray mutex, config/log path substitution, Run-key quoting
+  attack, startup config invalidity, network unavailable at boot, tray absent/
+  crashed, Explorer absent, service start timeout, forced stop with active Jobs,
+  uninstall interruption, log disk-full/rotation sharing violation, and Server
+  Core without GUI. Keep service live/not-ready where designed and never let tray
+  failure own daemon state.
+- `BH-SEC-01..22` — Exercise command/script path traversal, SQL metacharacters,
+  log/terminal escape injection, Unicode confusables in IDs, environment-value
+  redaction, secret-looking payloads in metrics/logs/audits, malicious PATH/
+  COMSPEC/PATHEXT/file association, unsafe executable replacement, named-pipe ACL
+  bypass, remote pipe access, inherited-handle leakage, decompression allocation
+  abuse, cursor forgery, JSON-RPC bind warning, unauthenticated JSON-RPC authority
+  as the explicit v1 behavior, and least-access state/work/log ACLs. Tests assert
+  the documented insecure boundary rather than pretending authentication exists.
+- `BH-OPS-01..20` — Handle asynchronous recovery timeout, one dirty scope, metrics
+  scrape during churn, log sink failure, full audit budget, invalid health path,
+  shutdown deadline, kill during shutdown, restart loop, incompatible/failed
+  upgrade, downgrade attempt, backup destination full, restore with wrong config,
+  clock rollback/forward/DST, hostname/client-ID change, and missing OS diagnostic
+  data. Liveness stays meaningful, readiness stays conservative, and absence is
+  never fabricated as zero/healthy data.
+- `BH-HARNESS-01..26` — Reject invalid/traversal run ID, corrupt/truncated/foreign
+  manifest, commit/image/config/VM mismatch, live lock stealing, stale lock with
+  live owner, unlabelled or wrong-repository Docker resource, symlinked run root,
+  path outside `.test-runs`, Windows resource outside test root, secret in report,
+  oversized artifact, no disk, unavailable image/runner, port collision, lost
+  runner, interrupted setup/collect/cleanup, failed Docker/SCM removal, unsafe
+  resume, reuse attempting mutable-state copy, GC of resumable run, purge without
+  confirmation, and `--all` with unrelated resources. Fail closed and print an
+  exact recovery/cleanup instruction without invoking global prune.
+
+#### Obscure race, crash-window, and interleaving cases
+
+All race cases run under deterministic schedule hooks first, then selected cases
+run repeatedly with the Go race detector or native process concurrency. A case
+must assert durable state, process count/identity, quota counters, emitted event
+order, and leaked goroutine/handle/file counts—not only the CLI exit status.
+
+- `RC-DOM-01..18` — Concurrent identical/conflicting command creation; mutation
+  retry while original commits; cancel versus queue dispatch, acceptance,
+  preparation, authorization, running, and terminal commit; signal versus cancel/
+  terminal; stdin write versus close/terminal; late resource snapshot versus
+  terminal; duplicate terminal events; command revision increments from multiple
+  controllers; UUIDv7 generation in one millisecond and across clock rollback.
+- `RC-SES-01..20` — Same-instance reconnects cross; old read/write loops race
+  fencing; different-instance reconnect races grant creation/expiry/consumption;
+  session closes while dispatch reserves capacity; capacity update crosses
+  acceptance/rejection; reconnect crosses queue expiry; stale output arrives
+  before/after new generation commit; reconciliation result is lost while new
+  dispatch wakes; two clients contend for global fairness; heartbeat timeout
+  crosses Pong/read activity/write deadline/server shutdown.
+- `RC-CLIENT-01..22` — Duplicate dispatch crosses durable local admission;
+  acceptance acknowledgement crosses service crash; queue dequeue crosses cancel/
+  shutdown; output/event sequence assignment crosses Ack/reconnect/compaction;
+  terminal acknowledgement crosses local cleanup/tombstone insertion; server
+  discard result crosses retransmission; local quota reservation crosses script/
+  stdin/output writes; reconnect crosses config restart; identity/spool recovery
+  crosses connection startup; two commands become runnable as capacity changes;
+  stop-all crosses a new prepared Job. Preserve at-most-once execution, monotonic
+  event order, and accurate advertised capacity.
+- `RC-STORE-01..28` — Event append races Ack/query/retention; terminal commit
+  races output drain and whole-command eviction; eviction races pagination,
+  follow, idempotency retry, and age rotation; tombstone insertion races full-
+  record deletion/replay; audit rotation races incident resolution; repair races
+  startup recovery/retention/backup; command/client/global reservations cross
+  exact caps concurrently; closeout consumes its reserve while disk floor trips;
+  group commit crosses process kill; WAL checkpoint crosses readers/shutdown;
+  two failures create incidents for one scope without reopening resolved history.
+- `RC-OUT-01..20` — stdout/stderr readers race root exit/descendant exit; child
+  inherits pipe while grace expires; raw backlog crosses high/low repeatedly;
+  compressor completion reorders across commands but not within one command;
+  Ack arrives during reconnect/spool compaction; output assignment races client
+  loss conversion; server retention marker races live follow; terminal event
+  races final output/incomplete marker; slow follower disconnects while page read
+  crosses segment eviction.
+- `RC-SCRIPT-01..12` — Duplicate chunks arrive concurrently; commit crosses last
+  fsync, cancellation, quota eviction pressure, client restart, digest failure,
+  and launch eligibility; cleanup crosses terminal acknowledgement/service death;
+  wrapper identity is swapped between validation and creation/revalidation;
+  identical upload resumes from competing old/new sessions.
+- `RC-WINCTX-01..18` — User logs on/off, locks/unlocks, disconnects/reconnects,
+  or changes console session during enumeration/token construction/revalidation;
+  session ID is reused with a different logon SID; multiple RDP sessions become
+  active/ambiguous; linked-token policy changes; profile unload races Job exit;
+  active-user CWD disappears or ACL changes before release. Re-enumerate at most
+  once as specified, never retarget silently, and never fall back after
+  `launch_prepared`.
+- `RC-LAUNCH-01..30` — Cancellation/service stop/client crash at every instruction
+  boundary among durable acceptance, token selection, Job/pipe creation,
+  suspended launcher creation, Job assignment, pipe authentication, suspended
+  shell report, prepared fsync, authorized fsync, release send, resume, running
+  acknowledgement, root exit, drain, terminal fsync, EventAck, cleanup, and
+  tombstone insertion. Duplicate dispatch after each restart must yield zero or
+  one process, never two.
+- `RC-SIGNAL-01..16` — TERM/KILL/cancel collide with launcher connection, shell
+  resume, root exit, descendant-only Job, helper pipe authentication,
+  `AttachConsole`, CTRL_BREAK delivery, grace timeout, Job termination, service
+  shutdown, PID reuse, and repeated signal revision. Record the actual attempted/
+  escalated result and never signal an unrelated PID/console.
+- `RC-SVC-01..20` — SCM start/stop/restart/uninstall cross recovery, command
+  admission, prepared launch, and active Jobs; two installers/configurators race;
+  tray startup crosses logon/logoff/Explorer restart/service restart; multiple
+  trays contend for the per-session mutex; pipe disconnect crosses a mutating
+  helper result; log rotation crosses tray open/read; config edit crosses explicit
+  restart; Automatic/Manual change crosses SCM query.
+- `RC-OPS-01..14` — Health/metrics/read RPCs cross recovery state transitions,
+  dirty resolution, retention, and shutdown; log/audit rotation crosses process
+  crash; backup crosses group commit and dispatch pause; upgrade crosses queued/
+  active commands and rollback; wall-clock changes cross queue TTL/retention while
+  monotonic heartbeat remains correct.
+- `RC-HARNESS-01..18` — Controller dies before/after manifest fsync, resource
+  creation/label recording, Windows lease, fault checkpoint, result write,
+  artifact collection, stop/reset/purge, and lock release; two cleanup commands
+  race; status/collect runs during cleanup; resume crosses a late old controller;
+  GC crosses a newly resumed run. Recovery must converge idempotently and never
+  delete an unrelated resource.
+
+For each crash-window family, implement a loop that enumerates named hooks rather
+than hand-selecting a few attractive points. Persist the hook name before
+triggering death, restart from the real store/spool, replay the same external
+request, and check the invariant matrix. A newly added durable write or external
+side effect must add a hook and coverage row before merging.
+
+### 2.6 Native Windows test-host requirements
+
+Native Windows is mandatory for the Phase 4/5 gates. Development can begin with
+unit tests and cross-compilation before a host is connected, but the Windows
+supervisor/service/tray implementation cannot be called complete without it.
+Prefer disposable or snapshot-resettable VMs over a personal workstation.
+
+Provision at least one primary interactive host with:
+
+- current Windows 11 Pro/Enterprise, Desktop Experience, Explorer, `cmd.exe`,
+  Windows PowerShell 5.1, UAC enabled, and all current updates captured in a
+  named clean snapshot;
+- at least 4 vCPU, 8 GiB RAM, and 60 GiB free disk recommended (2 vCPU, 4 GiB,
+  and 30 GiB free is the minimum smoke lane), with reboot and snapshot-revert
+  authority;
+- one local standard account and one traditional split-token local administrator,
+  plus a separate automation principal able to install/control the test service;
+- a preconfigured CI runner, OpenSSH, or WinRM management channel reachable only
+  from the test controller; credentials live outside the repository and run
+  manifest;
+- bidirectional reachability to the Linux nginx test endpoint, stable DNS or a
+  supplied address, time synchronization, and permission to transfer the
+  CI-built binary/config/CA into a dedicated test root;
+- no valuable user data, credentials, mapped drives, or production services,
+  because tests deliberately change SCM/Run-key state, ACLs, UAC fixtures, logon
+  sessions, kill processes, fill bounded scratch storage, reboot, and revert.
+
+Maintain snapshot/policy variants for logged-out, standard-user active,
+split-token-admin active, UAC disabled/already-full admin, and Administrator
+Protection when supported. The harness must restore the baseline after any
+variant that changes machine policy. A Windows Server Desktop Experience VM with
+multiple simultaneous `WTSActive` sessions is required for the final native
+ambiguous-session case; until available, keep its exhaustive selector unit test
+mandatory and mark only that native case blocked. Add a Server Core VM for the
+headless Session 0 release smoke. Before publishing, also exercise the oldest
+supported Windows 10 or Server 2016 baseline; it need not be the everyday runner.
+
+A physical Windows machine is optional. It is useful for an additional real
+display/audio/DDC command smoke, but RVBox only guarantees correct token/session/
+process execution—not success of arbitrary vendor hardware APIs—so physical
+hardware is not a release blocker. `test-host.ps1 Prepare` must inventory the
+host against this checklist and refuse destructive suites unless the machine is
+explicitly marked disposable/resettable and the clean snapshot identity is
+recorded.
+
 ## 3. Repository and build bootstrap (Phase 0)
 
 ### 3.1 Establish the repository layout
@@ -384,6 +805,7 @@ scripts/
   test-env             # doctor/recover/reuse/cleanup by exact run ID
   windows/test-host.ps1 # native Windows host lifecycle adapter
 test/
+  coverage.toml       # requirement-to-case inventory with stable IDs
   harness/             # shared manifest, journal, orchestration, and reporting
     defaults.toml      # local resource/time/artifact budgets
   integration/         # cross-package/real-resource suite definitions
@@ -425,8 +847,8 @@ Add:
   resource limits, run-ID labels, ephemeral PKI, fault proxy, and artifact/state
   mounts. It accepts only values produced by the validated harness manifest and
   never uses an implicit default project name.
-- `Makefile`: `generate`, `fmt`, `lint`, `test`, `test-race`, `test-integration`,
-  `test-e2e`, `test-status`, `build`, and `verify`. `test` aliases the unit
+- `Makefile`: `generate`, `fmt`, `lint`, `test`, `test-race`, `test-coverage`,
+  `test-integration`, `test-e2e`, `test-status`, `build`, and `verify`. `test` aliases the unit
   layer; integration/E2E targets delegate to the checked-in scripts and print
   their run ID. Each target invokes the container workflow and must not silently
   fall back to host tools.
@@ -451,11 +873,13 @@ Add CI (or a repository script ready for CI) that runs, in order:
    v1 compatibility baseline; enable `buf breaking` against that baseline
    immediately after it merges, not against the obsolete draft.
 2. Generation freshness check.
-3. `go fmt`, `go vet`, static analysis, and unit tests.
-4. Race tests for server/client concurrency packages.
-5. Linux server/storage/session integration tests in Compose; no Linux client
+3. Coverage-inventory validation: every required stable ID maps to a listed
+   case and every normative critical invariant has at least one case.
+4. `go fmt`, `go vet`, static analysis, and unit tests with per-package coverage.
+5. Race tests for server/client concurrency packages.
+6. Linux server/storage/session integration tests in Compose; no Linux client
    supervisor is required at this stage.
-6. Cross-compilation checks for Windows packages in the container plus native
+7. Cross-compilation checks for Windows packages in the container plus native
    Windows build/runtime smoke tests on a Windows runner. A Windows runner is a
    Phase 0 prerequisite, not a later optional enhancement.