Devostasis 0.1.8

Roadmap targets B5 and B1: the first accepted band ordering, and the format
and runner that make a conformance case executable. No threshold, window or
gauge changed, and no rule was retuned.

• Band ordering (target B5, PV-BAND-ORDER-001). Three Vitals now declare
  which of two of their own bands is better: Clutter
  CLEAN > LIGHT > CLUTTERED > HEAVY, Flow MOVING > CONGESTED > GRIDLOCKED
  for a live queue, Integrity CLEAN > FLAKY > FAILING and
  SPARSE > SPARSE_MIXED. delta.json therefore emits IMPROVED and
  WORSENED where the order applies. Pulse, Horizon, Direction and Debt
  declare no order at all, and neither does NO_QUEUE against a live queue or
  an Integrity evidence state against a verdict: those transitions stay
  CHANGED. No order is declared across Vitals or across projects, and none of
  this is a step towards an aggregate.
• A direction is only ever reported about exact measurements. The pair must
  be COMPARABLE, the rule_id unchanged and both sides AVAILABLE with
  EXACT band semantics; a DEGRADED band is a bound, and a move between
  bounds is not an improvement. The rule version boundary and observability
  transitions keep precedence, and a gauge that moved inside a band is still
  UNCHANGED. Every row that could have been ordered but was not says why:
  BAND_ORDER_NOT_DECLARED, BAND_ORDER_INCOMPARABLE or
  BAND_ORDER_NOT_ELIGIBLE.
• devostasis.delta.v2 carries the two new classes and names the ordering
  it applied in band_order_contract and band_order_version, so a stored
  delta says under which order its classes were decided. This is the fourth
  and, with the consumer surface now frozen, the last planned move of bundle
  identity: every row of the consumer surface in the ROADMAP is stable.
• Executable conformance vectors (target B1). A conformance case can now be
  written as JSON and executed: devostasis.vectors.v1 with a vital kind
  that evaluates one Vital over raw observation envelopes and a delta kind
  that compares two snapshots, a runner, a devostasis vectors command and a
  published schema (docs/spec/vectors.md). It fails
  closed: an unknown kind, an unknown key, an unknown comparison status, a
  malformed envelope or a duplicate case id is an error, never a skipped case
  that looks like a pass. The schema declares the partial evidence envelope a
  vector actually states, and a test holds it to the envelopes the corpus
  writes, so the runner and the published schema cannot accept different files.
• The ordering ships as fifteen vectors, not as prose. ORDER-01..15 live
  in tests/vectors/band-order.json, run inside the ordinary test suite and in
  CI through the command line, and the conformance table cites them as
  vector:ORDER-nn. A third drift guard now checks both directions: a citation
  without a vector fails, and a vector nobody cites fails.
• Those fifteen carry the meanings the contract gives them. The first
  version of this release derived the cases from the contract's rules instead
  of transcribing its Required conformance cases section, which moved every
  identifier from ORDER-02 onward onto a different claim — a research
  identifier is never reused, and conformance.md says so on the same page.
  PV-SPEC-001 found it; the cases are now transcribed, and the two accepted
  pairs the corpus never executed (NO_QUEUE against GRIDLOCKED,
  SPARSE_MIXED to FAILING) are executed. Semantics did not change: the
  ordering itself was conformant, and no band, threshold or rule moved.
  The split, adjacency and precedence cases this implementation wanted beyond
  the accepted set are kept under local DEV-ORDER ids, which belong to no
  research unit. A fourth guard holds each ORDER-nn to the accepted case name
  and to the band pairs that case names, because the first two guards pass
  happily while every identifier means something else.
• A vector case can state more than one pair. given.comparisons is a
  list, each entry with its own expect, and the case passes only when all of
  them do. Accepted cases are written that way — "GRIDLOCKED → CONGESTED → MOVING follows the WORSENED/IMPROVED direction" is six comparisons, "every
  unequal Pulse band pair" is twelve — and splitting one across several vectors
  would split its identifier. ORDER-01..15 now execute 85 comparisons between
  them. devostasis.vectors.v1 is unchanged for the single-pair shape it
  already had.
• The vectors of PV-TEST-001 are still owed by the research process. The 70
  named cases without a test remain open (debt D-1), but what was missing on
  our side is now built, so those vectors arrive executable instead of needing
  translation.
• A documented pin that names a tag nobody published is now a red build.
  tests/test_release_pins.py proves the pins agree with pyproject.toml; it
  cannot prove the tag they name exists, because on a release branch that tag
  legitimately does not exist yet. Nothing closed the window afterwards, and
  0.1.8 fell into it: every pin said v0.1.8 while master carried no such tag,
  so the install command in the README failed for anyone who ran it. A new
  Released pins resolve workflow runs daily on master and resolves every
  documented pin against the remote. Daily rather than per push, because the
  window is legitimate for as long as it takes to tag a merge and not a day
  longer. The self-observation workflow could never have caught this: it passes
  devostasis-ref: ${{ github.sha }}, which is right for its purpose and means
  the one live exercise of observe-self.yml overrides the input that goes
  stale (issue #18).
• Bookkeeping: the 0.1.7 section still said "unreleased" after v0.1.7 was
  tagged and released, which is the exact drift debt D-4 names.

Alongside the two targets, the defects the external review of 0.1.7 found
(#12, finding 5). Each was
reproduced before it was fixed. No contract, threshold or rule changed.

• A malformed title no longer ends a fleet run. _title crashed with
  IndexError on a title that is only whitespace and with AttributeError on
  one of the wrong type. Neither is a RegisterError, so neither reached the
  collector's error boundary. _title is now total for provider payload — a
  commit message, change-request or issue title of the wrong type is a missing
  title — while a register title of the wrong type is a RegisterError and
  reaches the snapshot as ERROR / INVALID_REGISTER, which is what the
  register contract says it should be.
• Linkage evidence is the exception, and it is now declared. A title or
  body of the wrong type is a missing title everywhere except where the target
  marker lives: _target_refs joined the raw fields, so a non-string there
  raised TypeError under planning.source = file, and reading it as an
  absent link would report a project unlinked on evidence nobody could parse.
  Unreadable title, body or milestone is now LinkageEvidenceError, a
  CollectionError, so the project fails explicitly with its reason while the
  rest of the fleet is still observed. A change request with no marker is
  still simply unlinked, because absent evidence is a fact and unreadable
  evidence is not.
• A successful response that is not JSON is a declared provider failure.
  UrllibTransport.get raised json.JSONDecodeError out of every handler on
  an HTTP 200 with an unreadable body. It is now ApiFailure with
  MALFORMED_RESPONSE, so it becomes an observation status like every other
  provider failure.
• One project's failure costs one project. run_all had no boundary of its
  own, so anything run_project did not anticipate stopped every project
  queued behind it, skipped the entity-tag cache write and left the fleet index
  stale. Each project now fails on its own, keeping its reason and its
  unsuccessful outcome, so the run still exits non-zero.