Loki Mode v8.0.1

Patch release. Every change is a bug fix; no behavior a working v8.0.0 install
depends on was altered.

Trust core

• The Evidence Receipt attested to an understated diff. _git_diffstat
  computed the receipt's stat against _LOKI_ITER_START_SHA -- the
  per-ITERATION baseline, which is re-captured to current HEAD at the top of
  every iteration -- while presenting the result as run-level. A multi-iteration
  run therefore attested only to changes since its last iteration, and that
  count feeds diff_sha256, the integrity hash written on every run. Measured
  on a real 2-commit run: 850 insertions + 76 deletions reported where the run
  produced 1479 across 10 files. Now resolves the run-level baseline, falls back
  to the empty tree for greenfield runs, and records base_sha from the same
  baseline the diff used so a verifier recomputing it cannot disagree.
• The verifier rejected the greenfield baseline it is meant to accept.
  _rev_resolvable probed only <ref>^{commit}; the empty tree is a TREE, so
  genuine proofs failed with "base ref unresolvable". Fixed without loosening
  the forgery defense -- fabricated, all-zero, and garbage SHAs are all still
  rejected.

Reporting

Every gate decision examined during this work was a TRUE positive. The defects
were in how those correct decisions were reported.

• A greenfield run reported files_changed: 0 for work it had done. A build
  that produced 4 spec-required files and passed 28/28 of its own tests told the
  user nothing was changed. Now diffs against the empty tree when no baseline
  commit exists: the same artifact reports 9 files, 1300 insertions.
• PAUSED.md never said why it paused. It was static boilerplate. It now
  names the blocking gate, its failure count and threshold, the findings path,
  and what was built before the pause.
• Terminal outcomes gave no next step. Guidance was reachable only for
  complete|max_iterations AND files_changed == 0, so intervention,
  failed, stopped, force_stopped and inconclusive_spec_contradiction
  got nothing -- exactly the outcomes where a user is stuck. Each now names a
  concrete next action, and the guidance is independent of the file count.
• loki doctor hid the inline-image protocol when bun was absent. The
  Cockpit section printed "Render path" but not "Inline-image protocol", so a
  user could not tell whether their terminal was unsupported or simply never
  probed.

CI and portability

The Tests workflow had not started for four days: a secrets reference in a
step-level if: is rejected at parse time, so ZERO jobs were created and the
only symptom was "workflow file issue". Fixing that surfaced a backlog of
platform bugs, each of which reported a product failure for an environment
difference:

• tar tzf | grep -q under pipefail reported a missing tarball artifact that
  was present -- grep -q closes the pipe and GNU tar dies of SIGPIPE.
• GNU stat -f means "filesystem status" and exits 0, so a macOS-first
  stat -f ... || stat -c ... never fell through on Linux (5 suites).
• An apostrophe inside a quoted heredoc nested in $( ) makes macOS bash 3.2
  reject the whole file, breaking both macOS Bun jobs.
• Three suites grepped a hardcoded /Users/<name>/git/loki-mode checkout, and
  two pinned /opt/homebrew/bin/python3.12.
• The shell-tests job installed a hand-maintained dependency subset missing
  sqlalchemy, failing five suites on one import; it now installs from
  requirements-test.txt.
• import mcp resolved to this repo's own mcp/ directory rather than the SDK,
  so a preflight guard passed with no SDK installed at all.

New guards prevent each class from returning: bash 3.2 parse compatibility,
stat ordering, hardcoded home and toolchain paths, greenfield diff-stat,
per-outcome guidance, and the receipt's run-level baseline. Every guard was
verified non-vacuous by reintroducing the bug it exists to catch.