Loki Mode

Faster convergence: the completion council evaluates the moment work is claimed done

Measured from real build telemetry (docs/SPEED-DIAGNOSIS-2026-07-01.md): a build could be
verifiably done at iteration 1 (code reviews non-blocking, tests green, checklist complete)
yet keep running needless "next improvement" iterations, because the completion council was
only allowed to evaluate every COUNCIL_CHECK_INTERVAL (default 5) iterations. One observed
build ran 14 iterations for a job finished in ~1, growing from 19 to 26 tests it never needed.

• An explicit completion claim now triggers an immediate council evaluation instead of
  deferring to the 5-iteration interval. When the agent invokes loki_complete_task (or the
  COMPLETION_REQUESTED fallback signal), council_should_stop evaluates on that iteration.
  This is a SPEED win that INCREASES trust alignment: the loop stops as soon as work is
  verified-done, rather than grinding past it.

• No verification was weakened. The fix changes only WHEN the council evaluates, never
  WHETHER it must approve. COUNCIL_MIN_ITERATIONS (a claim before the minimum still does not
  stop), the hard checklist gate, the verified-completion evidence gate, the aggregate vote,
  and the devil's-advocate re-review all run unchanged. A premature or unearned claim is still
  rejected and the loop continues -- never a forced green. The normal N-iteration cadence is
  unchanged when no claim is present.

• The claim signal is the STRUCTURED one (loki_complete_task / .loki/signals/ COMPLETION_REQUESTED), not the fuzzy log-grep the circuit breaker uses, so it fires only on
  a real, intentional completion claim.

• Regression coverage: tests/test-council-convergence-on-claim.sh extracts the real
  council_should_stop and drives the full truth table (claim off-interval evaluates now;
  no-claim cadence preserved; claim before MIN_ITERATIONS does NOT stop; file-signalled claim
  path). Proven RED on the pre-fix code, GREEN on the fix.

Honest scope note: this fixes convergence CORRECTNESS (the loop stops at verified-done),
proven by unit test. End-to-end wall-clock before/after on a fixed spec is tracked separately
(benchmarks/speed-benchmark.sh) and depends on a surviving long-build run; the fat-iteration
cost (a couple of disproportionately long iterations) is a distinct, still-open investigation.