Loki Mode

Adoption-friction + accuracy batch (ranks 6, 7, 10; RED/GREEN measured)

Driven from docs/research-2026-07/gap-analysis-backlog.json. Built in isolated
worktrees, each proven with a fixture RED/GREEN test, wired into run-all-tests.sh
AND local-ci.sh. All three are bash/python-only (no loki-ts mirror exists for
loki verify or the proof generator), so no Bun-parity surface is touched.

• rank 10 -- code-scope / locality record (accuracy, ADVISORY-first).
  loki verify now emits a structured scope record in evidence.json:
  {files_changed, net_lines, verdict, advisory:true, thresholds}. Verdict is
  "ok" under soft thresholds, "warn" over them, and "block" ONLY when an operator
  explicitly sets a hard cap (LOKI_SCOPE_MAX_FILES / LOKI_SCOPE_MAX_NET_LINES)
  and it is exceeded. Critically, the record NEVER feeds verify_compute_verdict and
  never emits a gate, so by default it CANNOT change the exit code or block a build
  (zero friction regression, verified). Greenfield builds are byte-identical by
  construction (the helper is reachable only via loki verify, never loki start).
  Test: tests/test-verify-scope-record.sh (4/4).
• rank 7 -- deterministic setup-recipe writer (adoption-friction). After a
  verified runtime boot, loki verify now writes .loki/setup-recipe.json
  {install, seed, env_keys, start, port, health_path} so later verify/e2e runs
  replay the exact steps instead of re-detecting heuristically (Devin's saved-setup
  pattern; the CONSUMER already existed). SECURITY: env_keys are variable NAMES
  ONLY -- the writer never reads the value side of any env line, proven by a test
  that plants a secret value and asserts it is absent from the recipe. A failed
  boot writes no recipe. Test: tests/test-verify-setup-recipe.sh (3/3, incl. the
  secret-leak assertion).
• rank 6 -- work-based engineering-hours estimate in proof.json
  (adoption-friction). proof.json gains a top-level effort_estimate
  {hours, low, high, method, model, inputs_hash, calibrated, label,
  estimator_version}. It is a work-based estimate (diff insertions/deletions +
  files + tests + brief size), so two builds of different scope get materially
  different hours (a 1-line tweak vs a full-stack feature differ by >3x -- the old
  flat "iterations x 15min" gave both the same 30min). HONESTY: it FAILS OPEN to
  the deterministic heuristic when no model is available (method="heuristic",
  model=""), and method="llm" ONLY when an opt-in LOKI_EFFORT_LLM=1 model call
  actually returned a parseable number -- it never fabricates an LLM figure. Every
  estimate is labeled calibrated:false ("uncalibrated") until a real
  ground-truth calibration dataset exists (tracked follow-up). effort_estimate is
  a TOP-LEVEL key, kept out of the facts block, so the LLM opinion can never leak
  into the deterministic proof headline. inputs_hash is a sha256 over the canonical
  inputs (recomputable from the written receipt) for auditability. The existing
  loki metrics iterations-based report is unchanged. Test:
  tests/test_effort_estimate.py (7/7); no proof-verify change needed (re-hashes the
  whole dict; 50 proof pytests still pass).