Loki Mode

Added

• Verified-completion evidence gate (default-on, opt out with
  LOKI_EVIDENCE_GATE=0). Loki no longer accepts a "done" claim without real
  evidence: the completion council blocks completion unless there is a nonzero
  git diff vs the run-start SHA (something was actually shipped) AND tests are
  green. This kills fabricated completion. The gate runs on BOTH completion
  paths: the interval-gated council path AND the default completion-promise
  route (an agent invoking loki_complete_task or emitting the promise text),
  so a self-asserted "done" with an empty diff or red tests cannot slip through.
  Diff evidence is the union of committed, staged, unstaged, and untracked work
  (--exclude-standard, so gitignored build artifacts do not count), with
  Loki's own .loki/ runtime state excluded so it is never mistaken for shipped
  work. Inconclusive cases (no git repo, no baseline, no test-results file,
  runner=none) pass through and never false-block a legitimate first run. When
  the gate blocks, it prints the reason and the LOKI_EVIDENCE_GATE=0 opt-out
  to the terminal AND surfaces it in the dashboard (/api/council/gate now
  reports an evidence block; the Quality Gates panel shows a banner).
• Bounded by design: a persistent block keeps iterating only up to
  MAX_ITERATIONS, then stops cleanly. It can never hang.

Fixed

• Stale-baseline bug on run 2+: a fresh loki start after a terminal run
  (success, failure, or crash) now resets the iteration baseline so the
  evidence gate diffs against the new run's HEAD, not the prior run's start SHA
  (which would have made the gate toothless). Genuine resume states (paused,
  interrupted, budget_exceeded, stopped) are deliberately preserved.
• Removed two SC2034 dead variables in the gate (the linter was red); the block
  decision is unchanged.

NOT tested in this release (honest disclosure)

• The gate proves "something changed and tests pass", NOT PRD-semantic
  correctness. A diff that compiles and is green but does not satisfy the spec
  still passes the gate; the council vote is the semantic check, not this gate.
• tests_red fires whenever a test runner ran and reported failure, including a
  project that was ALREADY red before this run. That is the expected
  false-block vector; the one-step LOKI_EVIDENCE_GATE=0 opt-out at the block
  site is the escape hatch.
• No live end-to-end autonomous run was used to validate the gate; verification
  is via the gate function against per-case throwaway git repos (33 assertions),
  the two-run lifecycle test (16 assertions), and the completion-route test
  (4 cases), all wired into the test suite.
• The Bun route is unchanged for this gate (the completion council is
  bash-only); this is a bash-route feature.