0.27.0

What a result is, one clock for a run, and an exact LP that closes.
No new commands (59) or kinds (57); the schema stays 5, with one optional
field on every kind (card).

New

• The card. A spec declares its result's role (consumer, existence,
  refutation, finite_check), the link of the chain it belongs to, and
  what it is still PENDING on: s.role(...).link(...).pending("CP7"), on
  every spec type, or "card": {...} in a JSON spec. The certificate
  carries it (optional field card, in the digest); verify prints the
  card -- kind, degree, role and link, scope, hypotheses used and cited,
  what is NOT established here, and the provenance (producing and verifying
  versions, schema, id; interpreter and package in --json, the MCP and the
  API). All of it labelled DECLARED: nothing about it is checked.
• Aggregates that say how much is checked. status counts the degrees
  (with --verify), the declared roles, and how many certificates rest on
  something pending -- a PASS there is conditional -- and lists pending
  dependencies and cited hypotheses with what is still owed. --oneline
  adds a seventh field, pending=k, only when there is something pending.
• The regime that matters. s.regime("l_ge_6", l >= 6, integers=["l"])
  declares where the result is meant, which is not a hypothesis; prove
  reports for each regime whether the hypotheses are inhabited there or
  EMPTY (the theorem vacuous where it was meant). check --hypotheses-only --regime NAME, and --integers.
• One clock for a run. --deadline (and each MCP call) sets a run
  deadline that every solver call, the engines with clocks of their own
  (ideal, the exact LP, profile, CEGIS), sweeps and branch and bound
  read: a run stops with TIMEOUT and what it had -- a sweep with the items
  not reached named, branch and bound with its frontier certificate --
  instead of being killed. Children (geng, cbc, lake) end with certo.
  The hard stop stays as the backstop. --timeout-ms keeps its meaning.
• The MCP server under a watchdog. Each tool call has a run deadline
  (CERTO_MCP_CALL_S, 600 s); a call that does not come back is answered
  without it, and past two such calls still running the server refuses new
  work until they end (CircuitOpen).
• cliques_of: every clique of a graph as a candidate, for "no
  partition into k cliques" (examples/clique_partition.py).
• FLINT for large polynomial products (fmpq_mpoly), exact, with the
  Python loop as the fallback (CERTO_NO_FLINT=1): the parametric examples
  run in 8.2 s instead of 11.7.

Fixed

• The exact LP on a large degenerate quotient closes. A 4130-column
  program never certified in a two-minute budget; the float dual is basic,
  and its basis is now solved exactly (denominators past the ladder, 2598216
  here), the primal from the float support: 3.5 s, verified complete. The
  exact simplex solves a wide program as its primal, pivots sparsely, and its
  phase 2 keeps the deadline (it ran past budgets).
• A Farkas core with an equality failed its own self-check. An equality
  is two rows and one name; requiring a name per row made prove report
  INVALID for a correct proof.
• A check model was labelled with a CNF witness line whose {n} and
  {names} were never filled.
• Branch and bound stopped by the run's deadline before its first design
  says so, instead of blaming the residual LP.