0.1.3-rc1

Added

• crew up and the installer report a below-floor box or rig, in the same
  words, naming crew's version and both platform versions found and wanted.
  It reports and never refuses: each call site keeps its own consequence,
  decided by the capability rather than the version (#679).
• The rig half reads only the boxes in the roster. A box on this host that
  crew was never asked about is left alone — as crew up and
  crew upgrade --all already leave it — so it is neither opened nor named
  in a finding (#679).
• The host schedule carries no weekly crew reset --all line, deliberately
  rather than by omission: the restore it performs takes duty.log with it and
  clears the resume breaker with no push, so it is deferred to 0.1.4 behind
  #328 (#678).
• crew reset itself is unchanged and available on demand — the verb, its
  version interlock, its refusals, the host lock and the job log all ship as
  they were, and only the schedule is affected (#678).
• shared/host-crontab.example argues the deferral where a reader who came to
  add a second line will look, and does not leave the line commented out: a
  commented job line is reinstatable by uncommenting, in the one file nobody
  diffs (#678).
• shared/docs/host-maintenance.md keeps the price and period reasoning the
  deferral is reversed from, and names what the deferral costs on disk until
  0.1.4 with the reading an operator should take (#678, #606).
• Kimi sessions resume once after eligible timeouts and report only resumed-turn usage (#674).
• Box vitals flag low disk or memory headroom from operator-overridable limits. (#613).
• The fleet floor shows each box's current/max memory and disk headroom. (#613).
• Session history rows open their available transcript tails in the floor overlay. (#612).
• duty.log now records that the box restarted, on the first tick after a
  reboot, so a session that vanished with the box has a stated cause on the
  line above it instead of an unbroken run of ticks. A box that has never
  ticked says so rather than claiming a restart it did not take (#609).
• shared/host-crontab.example schedules the fleet host's maintenance: crew restart --all daily, one line per job, with no flock and no redirect in the line (#590).
• crew restart and crew reset take one non-blocking host maintenance lock between them, so the two verbs can never overlap; the later run names the holder, touches no box, and exits 4 (#590).
• crew restart and crew reset write every run to a host job log, boundary lines and per-box outcomes alike, so a run that did nothing still says so and silence at a boundary means cron is dead (#590).
• CREW_HOST_STATE_DIR moves the host maintenance lock and job log together (#590).
• shared/docs/host-maintenance.md is the runbook for the host jobs: what a skipped box means, what a refused reset means, and the rule that after any crew upgrade you re-cut before the next reset (#590).
• crew reset --cut snapshots each box as armed after proving it is logged in — GitHub and the agent profile's vendor CLI — and hired, reclaiming with an on-demand reaper run before it measures the disk and refusing a box still over the ceiling by naming what is large (#589).
• crew reset restores each box to its armed checkpoint and starts it, refusing a box with no such label rather than falling back to another one (#589).
• crew reset refuses a checkpoint whose recorded engine version is missing, unreadable or not a version, and --cut refuses a box whose engine stamp it cannot parse (#589).
• --force-after forces the restore path only; a --cut is skipped on a held duty lock however old (#589).
• crew restart drains and cycles selected boxes, skips live sessions, and reclaims stale duty snapshots (#588).
• Operators can price captured token records from an editable per-model rate card without stamping mutable costs into session history. (#572).
• kimi sessions record their token usage and credential pool on SESSION END (#571).
• A multi-step kimi session's totals cover every step rather than a single one (#571).
• A kimi session records uncached input, output, cache reads and cache creation, and no cost (#571).
• A kimi session's usage names its model unknown (#571).
• A kimi session whose usage is unavailable or unreadable records none, and its result is unchanged (#571).
• Codex sessions now record uncached input, cache, output, session identity and honest unknown-model usage while preserving final prose and omitting unavailable cost. (#570).
• Codex session logs preserve intermediate messages, command output and completion status when structured usage capture is active. (#570).
• SESSION START and SESSION END carry sid=, the session id the engine mints and pins on the CLI, so a session's transcript is addressable from duty.log (#538).
• A session killed at its wall is continued once rather than restarted, where its head has not moved, its log is non-empty and nothing of it survived (#538).
• Review sessions can park on bounded detached checks and resume once with recorded results instead of restarting long verification on every tick. (#533).
• SESSION END records log bytes and surviving session processes (#529).
• A release candidate carries its own drill record, drills/X.Y.Z-rcN.md, and
  drills/README.md states the ladder the release doors already implement
  (#506).
• A release guard refuses a final tree whose diff over its last candidate's tag
  touches anything but VERSION, CHANGELOG.md, changelog.d/ and drills/
  (#506).
• Deleting a candidate's drill record cannot lower the anchor a final is
  measured against: the rc-ladder guard reads the published candidate tags as
  well as the shipped tree's records (#506).
• The rc-ladder guard requires every reachable published candidate's own record,
  spelled exactly as its tag, so renaming a record under another spelling of its
  number is refused as the deletion it is (#506).
• A zero-padded candidate X.Y.Z-rc0N is measured as rung N rather than
  reported as a tag that cannot be resolved: the ladder is ordered by the number
  and named by the spelling (#506).
• Only the release doors' own spelling is a candidate — X.Y.Z-rcN in ASCII
  digits — so a version or tag the doors would never classify never anchors, and
  never gates, a release (#506).
• A stray space inside VERSION cannot hide a release window from the rc-ladder
  guard: the version is consumed the way the release doors consume it, with all
  whitespace deleted rather than only the ends trimmed (#506).
• The rc-ladder guard reports a shallow clone as a could-not-look, including
  when ancestry is what it cannot decide, and names the fetch-depth: 0 fix
  (#506).
• Fork PR authors can flag an evidenced red check for privileged rerun servicing. (#505).
• Triage sessions carry the board's round count for every issue, with the
  distribution those counts form and the round cap they are read against, as
  sizing evidence at the mint (#503).
• The builder engine records every open authored pull request's round count
  against the issue its body names, not only the ones that have reached the cap
  (#503).
• The builder engine counts a PR's review rounds and names the round cap to the
  session when the fifth round closes without a full approval standing at the
  current head (#502).
• CI refuses a tracked path that no scope:* row in .github/labeler.yml matches, naming each one; a path deliberately left unmapped is declared with its reason in .github/scope-coverage.allow (#500).
• Writing a scope:* label twice in .github/labeler.yml is refused, naming both lines: the scope job keeps only the last block, so the earlier one's globs derive nothing (#500).
• The release records the installer's size and payload file count, on the build's output and in the workflow run summary (#499).
• The rehearsal runbook documents every leg of the drill round — what each one
  needs, what it produces, and what its failure means — and names the blockers
  that stop a leg running on a bare host (#497).
• CI diffs the runbook's documented leg set against the harness's own
  declaration in both directions, so a leg added to the round cannot go
  undocumented (#497).
• The teardown fixtures drive the notifier union leg's second sandbox per role
  through the delete set, the ownership refusal and prefix-is-not-membership
  (#496).
• The fleet floor edits both watch registries — the fleet-wide repos.txt and
  notify-repos.txt, and a per-box selection reached from that box's own
  console (#488).
• A box's registry selection is repos.d/<box>.txt beside the fleet-wide list,
  and crew hire and crew upgrade transport it in place of the fleet-wide
  one (#488).
• A box selects from the fleet-wide registry and can never reach outside it: an
  entry the fleet does not watch is refused at the edit, and one retired
  fleet-wide leaves every box that had selected it (#488).
• A registry edit naming a malformed entry, or a repository the fleet cannot
  reach, is refused at the edit with the reason (#488).
• Every registry write is appended to .registry-journal.log in the fleet
  definition, naming the operator, the file, the state that write applied and
  the entries that moved, so one line describes the registry it produced
  without reference to any other (#488).
• Registry writes are serialised per registry and written through a temporary
  file of their own, so two edits in flight leave the file at one of the
  submitted states and one accurate record each (#488).
• A registry write that cannot be recorded is refused before anything moves,
  one the filesystem will not accept is refused with the reason, and one whose
  record fails after the file changed says the edit landed unrecorded rather
  than reporting a refusal (#488).
• crew reads every registry file through one reader, so a hand-indented entry
  is the entry it names and the console and the transport resolve one file
  alike (#488).
• A fleet-wide registry the definition does not have is shown as served from
  the shipped examples/ file the transport stages, and saving it writes the
  definition's own copy (#488).
• Force stop as its own floor action, killing a guest that cannot schedule
  its own shutdown (#486).
• crew status and the fleet floor show bounded per-box tick health from existing duty records (#484).
• Every tick emits a box vitals record — cores and load, memory total/available/shared, swap configured and active, disk total/used/percent, platform and OS release (#483).
• A vitals reading that disagrees with the role profile's BOX_CPU, BOX_MEMORY or BOX_DISK is marked on the record as a finding, naming both the declared and the measured figure — over the declared figure as well as under it (#483).
• Swap configured and swap active are recorded as separate fields, so a swapfile present while swapon --show is empty reads as a finding rather than as free disk (#483).
• crew status <box> and the fleet floor both render the box vitals record, from the same logged line and in the same words, so the two consoles cannot describe one box differently (#483).
• The record's disk history is backfilled from boot-check.log, so the series starts where the box did rather than at the first tick after the probe shipped (#483).
• The duty engine declares its operating limits and reports near-limit and crossed-limit measurements with their affected repository and pull request (#482).
• The floor relays operating-limit events from boxes through its configured operator alert channel (#482).
• The fleet floor can send one configured alert when a roster box becomes unreachable and another when it recovers (#481).
• Builder resume suppression now appears with its age and reason in crew status and on the fleet floor (#480).
• Session-end records can include Claude token usage, actual-model attribution, conditional cost, session ID, and an operator-declared credential pool without changing prose logs. (#475).
• A session now runs under a memory ceiling: SESSION_MEM_MAX_PCT in
  fleet.defaults.conf, a percentage of the box's MemTotal, with a per-role
  override beside it. A session whose measured peak crosses it is terminated
  and its SESSION END carries outcome=MEMORY (#474).
• The termination covers the session's own process group as well as its pid
  tree, so a descendant that handles TERM by forking a replacement and
  exiting cannot outlive the session that was ended (#474).
• The ceiling ships at 0 on every row, which is off: with nothing configured
  the engine starts no watchdog and a session behaves exactly as it did (#474).
• SESSION END now carries peak_rss=, the largest VmHWM in the session's
  process tree in KiB, so a session that grew until the kernel killed something
  is visible in the engine's own record of it rather than only in the silence
  that followed (#473).
• The fleet floor renders that figure for the newest session per box, and shows
  a dash for a session it has no figure for — including every line written
  before this (#473).
• The field is omitted entirely where the engine got no reading: a platform
  that reports no VmHWM, or a session shorter than one measurement interval.
  An absent field is not a zero, and nothing an aggregate reads can confuse the
  two (#473).
• A session that died with its box gets peak_rss=- on the terminal the orphan
  reconciler writes for it, beside the rc=- and dur=- it already carried: a
  measurement owed and lost, not one never taken (#473).
• Fast triage ticks now wake when a declared predecessor changes state and name the dependency edge that needs a prose re-read. (#471).
• A duty may now name its own model tier: MODEL_<KIND> beside that kind's
  timeout in conf/roles/*.conf, and run_session invokes it instead of the
  agent profile's default, so a mention and an issue mint need no longer be
  bought at the same price (#469).
• The agent profile owns the translation into its own flag, and a profile that
  cannot express a tier leaves the duty on the default and warns, naming the
  profile and the kind, rather than ignoring the setting silently (#469).
• SESSION END now carries tier=, naming the resolved tier or default, so
  an aggregate can tell a duty that got cheaper from one that got rarer. The
  field is appended to every line, unconfigured sessions included (#469).
• Per-duty tiers ship unset everywhere: an engine that is upgraded and not
  configured invokes exactly what it invoked before, and every session behaves
  as it did. The tier=default field above is the one deliberate change to
  what a SESSION END line looks like (#469).
• The duty engine now enforces a session budget: a per-kind ceiling on both
  sessions and session-minutes over a rolling window, checked in run_session
  before any vendor call (#464).
• A lane at its ceiling logs SESSION SKIP … reason=budget and stops
  dispatching until the window rolls, leaving every other kind untouched
  (#464).
• The operator is told at 80% of a ceiling as well as at it, once per crossing
  per kind, on the channel the session breaker already uses (#464).
• Session budgets are configured per role and per kind in conf/roles/*.conf
  and ship OFF, so an engine that is upgraded and not configured behaves
  exactly as before (#464).
• crew status <box> reports each configured lane's balance — spent, ceiling
  and when the window next frees capacity — from the moment a ceiling is set
  rather than from that lane's first dispatch, and omits the section when no
  lane is configured (#464).
• An unreadable or unwritable budget counter refuses to dispatch and alerts,
  rather than dispatching on a balance the engine cannot read (#464).
• Triage prompts can route vendored-doctrine gaps to a configured, fleet-listed upstream repository (#463).
• A builder that judges a ready issue unbuildable now declines it on the
  board: a comment on that issue naming one of four reasons and @-mentioning
  triage, instead of a decision that died with the box (#462).
• post-once.sh takes an optional marker line and dedups on it, so a comment
  whose identity is narrower than its text posts once (#462).
• A daily reaper on every role reclaims what this engine's own sessions
  leave on a box: CLI transcripts older than the retention age, and cache
  entries nothing has read inside it. Each class logs its byte figure, and
  logs one when it reclaims nothing (#457).
• The reaper holds rather than deletes whenever it cannot establish a fact:
  while any session is live, on a filesystem whose access times do not
  advance, and over anything it could not finish reading (#457).

Changed

• The mint sequence has one writer that crew new and the drill both call,
  instead of two copies a human kept agreeing (#679).
• crew declares the platform it is built and tested against — box 0.10.0 and
  rig 0.4.0 — in one place, and the two call sites that hardcoded a version
  read it from there (#679).
• Codex duties resume one eligible timed-out thread instead of restarting from zero (#673).
• The crew status STUCK note names the ceiling that was exceeded, or says
  no session was in flight, matching the floor's wording (#624).
• Root can install one system crew tree that each host operator runs with their own identity and fleet configuration (#615).
• The STUCK note and panel name the ceiling that was exceeded, not only how
  long the lock has been held (#610).
• A freshly minted reviewer box is 4 CPU, 8GiB, 60GiB — the builder's size
  (#607).
• reviewer.conf's comment above the figures names what a review run does, and
  no longer calls the role lean or its worktrees throwaway (#607).
• A box cloned from a gold snapshot — crew new --from, or a roster line with a
  fourth column — is sized at the role's cpu and memory where the host's box
  supports it, instead of inheriting the snapshot's (#607).
• --disk is never passed on a clone, so a cloned box keeps the snapshot's disk
  (#607, box#171).
• Every clone ends in one line giving cpu, memory and disk each applied, NOT applied or unverified, beside the size the box carries, the size the role
  profile asks for, and the commands to resize all three (#607).
• drill/rehearsal.sh mints each drill box at its role's size (#607).
• The ceremony pin moves from 0.7.6 to 0.7.7 across all ten anchored
  uses: references — the six release guards, both label callers, the release
  door and the privileged rerun service — so a fresh taxonomy bootstrap no
  longer stops at the operator label's over-long description (#603).
• Reviewer duty coverage now runs in its own entrypoint-driven test suite. (#601).
• crew restart and crew reset exit 4 when nothing was attempted, distinguishing it from a run in which everything attempted worked (#590).
• crew hire, crew up and crew upgrade mark the box's armed checkpoint stale before installing an engine, and refuse to install when that mark cannot be written (#589).
• crew reset refuses a box whose armed checkpoint is marked stale, naming both versions and the re-cut (#589).
• crew hire and crew upgrade both name the stale mark they left when the install that followed it failed (#589).
• crew new --from refuses an armed checkpoint as a clone source (#589).
• crew down waits for duty locks, reports partial failures, and requires an explicit supported --force to stop wedged boxes (#588).
• Reviewer prompts now show same-head CI conclusions and qualify how that evidence should be used (#532).
• fleet-floor/server/floor.py is split into a floor package on the banners it already drew — ping, roster, units, fleet, actions, alerts, server. Both doors are unchanged: floor.py still starts it, import floor still answers with the whole collector (#508).
• fleet-floor/test/ mirrors the collector's modules: cases.sh becomes one suite per module under test/floor/ (#508).
• shared/lib/common.sh is now an entry point over seven subject modules under shared/lib/common/; callers source the same path and get the same functions (#507).
• Fixtures for those modules run as seven suites under shared/test/common/, one per module at the mirrored path (#507).
• Agent doctrine and release machinery now use ceremony 0.7.6. (#505).
• PR Round logs now keep bounded review facts and reply links instead of copying whole-round replies into the body (#504).
• The engine holds its panel request on a PR at the round cap, so the successor
  is the only PR the panel is asked for (#502).
• A fifth round that closes with the panel's unanimous approval, where those
  approvals still stand at the current head, converges and hands off as usual
  and is never named for the cut. Where the head has since moved the approvals
  are spent and the PR is at the cap again (#502).
• The published installer carries only what an install needs, instead of the whole tracked tree: crew-<version>.sh is 531 KB where it was 51 MB (#499).
• Restart force-stops an unreachable box before starting it; a box that still
  answers restarts gracefully (#486).
• Lifecycle action replies name which step each per-box row came from, and the
  status line says when an action force-stopped a box (#486).
• The fleet payload serves the collector's reachability verdict as
  ping.wedged, so the console reads one rule rather than re-deriving it from
  the miss count against a threshold of its own (#486).
• A restart carries the mode its confirmation named, and the floor refuses it —
  changing nothing — if the box crossed the reachability boundary in between,
  so a graceful confirmation can never authorise a force stop (#486).
• SESSION START carries holder=<pid>.<boot>, the duty process that
  dispatched the session, so a session still running is told from one that
  died by asking its holder rather than its age (#478).
• Shared duty-engine fixtures run as five standalone subject suites behind the aggregate entrypoint (#470).
• Fast-tick triage sessions now receive only the named issues and discussions that woke them; hourly hygiene retains the whole-board sweep (#468).
• The hourly hygiene sweep now launches a session only for repositories whose board changed since the last one, and sweeps a never-changing board anyway once every HYGIENE_FLOOR (#465).
• A hygiene gate that cannot read the board or its ledger sweeps anyway and warns by repository, a HYGIENE_FLOOR that is not a whole number of seconds falls back to the shipped default with a warning rather than disarming the floor, and every withheld repository is logged each interval (#465).
• The hygiene prompt no longer asks for the blocked-to-ready flip or the stale-claim reclaim, which the issueflow reconciler already owns (#465).
• no build duty names why the ready work is held —
  2 ready declined: unbuildable (1), needs-ruling (1) — where it named only
  the ledger holding it (#462).
• Builder-specific panel rows now enroll both live builder identities without changing required reviewers (#458, heavy-duty/ceremony#424).

Fixed

• The repair crew hire and crew status print for an unconverged box now opens box root rather than box shell plus sudo (#700).
• That repair now installs rig at the declared pin when the guest has none (#700).
• crew help reset now describes the daily restart schedule and on-demand resets without advertising a removed weekly job (#688).
• crew status reads the newest vitals record even when an older log segment contains binary data (#688).
• Reviewer sizing guidance now accounts for clean worktree reclamation while retaining dirty-worktree capacity (#688).
• crew new mints a box again on the box version crew's own floor requires:
  a blank mint plus a rig bootstrap <agent>-box, replacing the
  --template <agent>-box spelling box 0.10.0 retires (#679).
• The guest tenant keeps the agent's own name — claude, codex, grok,
  kimi — rather than silently becoming the blank seed's dev (#679).
• drill/rehearsal.sh no longer dies at box creation on a current box host,
  so the rehearsal can reach its assertions (#679).
• crew help, crew new's success response, examples/fleet.roster and
  shared/docs/single-role.md described a retired <agent>-box box
  template; all four now describe the blank mint and rig's <agent>-box
  role (#679).
• crew new's response no longer says the vendor CLI is converging when rig
  has already finished, and makes no rig claim at all on the clone path,
  where no bootstrap ran (#679).
• The platform report no longer reads box: not found found, and says the
  guests could not be read rather than reporting an empty fleet when jq
  is absent (#679).
• Verdict submissions keep their byte-identical retry copy on the duty volume, so a full volatile temporary filesystem cannot discard a completed review (#672).
• Reviewer sessions retry missing exact-head verdicts up to three times instead of suppressing live requests after a successful model turn (#671).
• Builder checkouts now validate and cache their writable fork, repair stale remotes, and refuse dispatch when no unique head repository exists (#635).
• Floor-only pull requests now run the shell suite, including its repo-wide
  and cross-reader guards (#625).
• crew status no longer reports a healthy box STUCK for spending time it
  was dispatched to spend: the lock age is measured against the ceiling the
  in-flight session declares, so the CLI and the floor grade one box alike
  (#624).
• The fleet god-view now flags usage-limited lanes with their cause and lane, and distinguishes terminal sessions from ordinary failures (#611).
• Unit consoles now show collected operating-limit events and retention losses even when no alert channel is configured (#611).
• The floor no longer reports a healthy box STUCK for spending time it was
  dispatched to spend: the lock age is measured against the ceiling the
  in-flight session declares, so a build entitled to 3600s is not alarmed
  at minute 10 (#610).
• Fleet probes read tick age and vitals past NUL bytes in duty.log, avoiding false SILENT boxes (#608).
• Reviewer ticks reclaim finished review scratch while preserving dirty work and verdict records (#606).
• Reviewer ticks reclaim finished review scratch without reading the PR's state (#606, #597).
• A dirty review worktree is preserved by moving it to kept-<name>-<head> rather than left on the path the next review needs (#606).
• Reviewer repo-wide checks now run from the detached PR checkout, excluding sibling worktrees and mutation fixtures. (#605).
• Session records now distinguish kernel OOM kills from wall-clock timeouts (#600).
• Productive memory-killed sessions resume with cause-aware guidance, while error-only logs start fresh (#600).
• Reviewers cap local ShellCheck memory, never retry kernel-killed checks, and rely on same-head CI for repo-wide lint. (#598).
• Review ticks reclaim detached worktrees left by ended sessions before dispatch, while preserving branch-holding builder worktrees (#597).
• Fleet-floor test call logs now keep each concurrent box invocation in one atomic record (#563).
• Reconstructed orphan session records now retain every observed field, using unknown markers when the original value cannot be recovered (#553).
• A terminal the vendor did not cause is no longer cleared by a vendor probe
  succeeding, so a box that kills its own sessions is stopped rather than
  throttled to three dispatches every four ticks (#551).
• A lane held by such a terminal clears on an observed session end instead —
  a session that ran to its own SESSION END, whatever its outcome, which is
  the evidence that the box survived it (#551).
• A held lane is retried on a widening interval bounded by
  SESSION_TERMINAL_HOLD_MAX_TICKS, so it is never tried every tick and never
  left permanently dark (#551).
• The breaker's trip alert fires on the transition alone, so a held lane's
  trial dispatches no longer raise an alert pair per cycle (#551).
• Operator-launched sessions now inherit the layered fleet and role configuration and use the longest applicable session timeout unless explicitly overridden (#537).
• Reviewer sessions now use bounded PID or sentinel waits instead of self-matching process polls (#530).
• Fleet-wide mention batches stay visible without becoming links to a repository named fleet (#528).
• crew upgrade from an installed tree no longer ships __pycache__/ directories into the payload (#517).
• Duty logs name missing session-classifier hooks once per agent profile per boot (#501).
• Claude quota-reset sessions now reach the terminal breaker (#501).
• A pull request touching only postmortems/, .box/README.md, shared/crontab.example, .github/workflows/labels-sweep.yml or the release artifact hook now carries a scope label (#500).
• Installer payloads now exclude ignored dependency trees and reject known-excluded paths before installation (#498).
• The runbook's Phase 3 report step describes the target the round derives from
  the ref it drilled, instead of routing findings to a long-merged pull
  request (#497).
• Drill records now enumerate every declared leg and explain each leg that did not execute. (#495).
• Drill rounds distinguish disarmed-only agreement from a measured armed, ticking, clock-skewed comparison and run both without extra boxes (#494).
• Drill rounds without an armed roster stay incomplete and retain their boxes for inspection (#494).
• Rehearsal reuse now refuses dirty sandboxes, while every role removes only the exact fixtures its current run created, on both success and failure (#493).
• The drill's exit footers name the pull request derived from the ref actually
  drilled, instead of a hard-coded long-merged one (#492).
• A drill round whose source names no pull request — a branch, a tag, a commit,
  a local --tree — prints no report instruction rather than a stale one (#492).
• Rehearsal rounds continue independent sections after a phase failure and report every section as passed, failed, or skipped-by in the summary (#491).
• Role rehearsals now resolve a candidate ref once and drill the same commit in every box (#490).
• The fleet-floor browser walk now checks byline overlap against one settled viewport frame at a time. (#489).
• crew hire reads a box with a narrowed registry as narrowed, rather than as
  a box about to be armed against the production registry (#488).
• crew hire compares the registry a box carries against the fleet's through
  the same reader on both sides, so a trailing space or a CRLF in either file
  no longer reads as a narrowing and lets an off-roster box be armed against
  the production registry (#51, #488).
• A floor restart reads its own stop result before starting: a stop that failed
  is reported as a failed restart and the start is not attempted, rather than a
  box being started by a lever that never stopped it. A box that was already
  stopped is started, not reported as a failed stop (#487).
• wake-silent no longer sends its wake into a box whose box exec is wedged.
  That box is named for a restart instead of costing the action a full timeout
  to report a wake it could not deliver (#487).
• Resume detection no longer loses a pull request whose comment thread grows
  past 131,072 bytes: the thread reaches jq down a pipe rather than as one
  argument, so the bound that applies is the one that scales (#479).
• Every failure on that path now warns (#479).
• A structural failure there — one that will not clear on its own, unlike a
  transient gh error, which is still retried next tick — now reaches the
  operator once per head instead of skipping the pull request forever (#479).
• That escalation also leaves a comment on the pull request naming the head, so
  the record survives an alert that was never delivered (#479).
• A session killed with its box no longer leaves a SESSION START nothing
  answers: the next tick emits the missing terminal with outcome=ORPHANED,
  and with rc=- and dur=- rather than a figure it cannot know (#478).
• A reconstructed terminal counts toward its kind's session breaker, on the
  same counter an observed one feeds: three in a row trip that lane and it is
  refused for the rest of the tick that trips it, until the next tick's CLI
  probe succeeds and clears the breaker (#478).
• Doctrine checkout warnings now report each frozen state once, include its age, and announce recovery without modifying session-owned Git state (#477).
• Credential failures now report rejected GitHub credentials and API responses instead of blaming configured git authors (#476).
• Git identity failures now name configured authors and the GitHub login in use (#476).
• The session subshell is raised in the kernel's OOM victim scoring before the
  CLI starts, so a box out of memory loses the session that is growing rather
  than cron, sshd or the engine itself (#474).
• The fleet floor's box probe lists *.log from the log directory rather than
  everything in it, so a poll landing while a session runs no longer spends one
  of its forty slots on that session's scratch file (#473).
• acted= in duty.log reports yes, no or unknown on claude and grok boxes, where it was always unknown (#467).
• Triage batches fresh mention threads from every repository into one capped session per tick (#466).
• Builder signals exclude ready issues marked as operator-owned work. (#461).
• Claims now stop before editing repositories where the claimed label is unavailable (#459).
• Failed issue claims now preserve the ready queue state and report repair or withdrawal failures (#459).
• The breaker drill reports profiles without bot_session_terminal as named INCOMPLETE skips instead of failures (#454).
• A maintainer's CHANGES_REQUESTED now wakes the builder to a fix round,
  once per block rather than forever (#452).
• The handoff no longer refires under a standing human change request, so a
  handed-off pull request stops bouncing back at the human with a fresh
  review request and a fresh notification (#452).
• A round answered with argument at an unchanged head still converges, so the
  argument reaches the human (#452).