devagent 1.0.0

First tagged release, cut at go-public (#463, #464). Everything in this section
had already shipped SHA-by-SHA to master before the tag existed.

• The plugin carries a version: plugin.json version: 1.0.0 (#532
  re-take). Behavior change for every install from the GitHub marketplace,
  stated here because #532 required it never be claimed as "no change": until
  now claude plugin update devagent@devagent followed master commit by
  commit; from this release it resolves to the version string and only sees an
  update when that string changes, so a merge to master reaches installed
  users at the next release rather than immediately. A marketplace added from
  a local checkout loads it in place and is not pinned. Measured on Claude Code
  2.1.223: claude plugin validate --strict now passes (it was
  documented-red on the missing version), and claude plugin tag --dry-run
  resolves devagent--v1.0.0. tests/test_plugin_versioning.py inverts from
  "no version anywhere" to "semver in plugin.json, none in the marketplace
  entry, matching CHANGELOG section", and tests/plugin-validate.bats wires
  the strict gate plus the tag dry-run. README, the docs-site install page and
  SECURITY.md describe the release policy instead of SHA-tracking.

• The onboarding site can be built and published (#465).
  scripts/build-docs-site.sh renders the docs-site/ pages to a static
  site with pandoc, straight from the tracked markdown — no rendered copy is
  committed. It rewrites page-to-page links, sends links that leave
  docs-site/ to the forge, and fails the build on a link it cannot classify
  or that resolves to nothing, and refuses a pandoc older than 3.x. A new
  publish docs site workflow runs the builder's tests, builds, and can deploy
  to GitHub Pages by keyless OIDC. Deploy is off until the repository
  variable DOCS_SITE_DEPLOY is set to true, because GitHub Pages is not
  enabled on the repository yet and a deploy against a Pages-disabled repo
  fails; docs/pages-deployment.md has the switch-on steps. New optional test
  prerequisite: pandoc 3.x, for tests/build-docs-site.bats only.

• The oneshot tier's no-repo-diff boundary is enforced mechanically (#595).
  "An operational action, not a repo change" lived in prose in three documents
  and nothing caught a one-shot that quietly produced commits. The oneshot tier
  has no branch step, so no baseline_sha exists and zero_diff_classify
  answered indeterminate forever. What a branchless tier measures is now
  defined as whether the source tree carries unpublished change — on the
  base branch, no commits beyond default_baseline, no uncommitted paths —
  rather than whether the issue itself produced a diff, which a shared source
  tree cannot attribute. A new scripts/oneshot-zerodiff.sh owns the
  predicate: it calls the shared zero_diff_classify unchanged for the commits
  half, adds a git status --porcelain probe for the dirty half (a dirty-only
  tree classifies as empty), compares the checked-out branch to the base, and
  mints one exit code per verdict. cleanup.sh reads the tier itself and
  refuses before any side effect on violated or indeterminate, directing
  the operator to revise.sh --retier standard; every other tier pays one
  header read and never spawns the checker. A shared tree left on a sibling
  issue's branch is the routine indeterminate, with a one-line remedy; a
  recorded linked worktree is indeterminate with a remedy that names the
  tier. The refusal never asserts authorship.
  echo "<reason>" > <issue-dir>/.devagent-oneshot-ack is an auditable escape
  seam for trees git cannot measure.
  commands/document.md's straggler instruction — which told a oneshot to
  commit and promised a step 12 that tier does not have — is carved out, and
  commands/implement.md's "must produce no repo diff" is reframed to the
  decidable form; both key on the tier name, since the research tier omits
  step 12 too.

• Untracked new source files enter the analyzer's changed-line scope (#591).
  static_analysis_diff.py scoped from git diff alone, which lists TRACKED
  changes only — a brand-new file that had never been git add-ed produced no
  hunks, so every per-file tool skipped it and step 13 passed vacuously for exactly
  the file with the least review history. A new get_untracked_ranges() enumerates
  untracked, non-ignored candidates with
  git ls-files --others --exclude-standard --full-name -z (read-only — no
  git add, no git add -N, no temp commit) and gives each a whole-file
  LineRange(1, <line count>), so cpplint, codespell, ruff, flake8, bandit, mypy
  and cmake-lint run on it and filter_novel counts every line as changed; markdown
  runs print one artifact-visible progress line,
  Untracked files (whole-file scope): … (stderr under --json, #119), and the
  summary labels now say "changed lines or untracked files". The compile-database
  tools (cppcheck, clang-tidy, iwyu, compiler warnings) and git clang-format
  cannot reach a file the build / index does not know and still report it clean —
  a documented limit. Candidates are limited to the suffixes the run_* filters
  already select on (.cc .c .h .py .cmake CMakeLists.txt) and exclude the
  analyzer's own build_dir / -asan / -ubsan / -tsan dirs and any root-level
  build-* dir — where analyze-sanitizers.sh builds regardless of the configured
  build_dir (#324/#351) — which a target project's .gitignore may not cover.
  Names are decoded as UTF-8 (never the locale's encoding), and a listed candidate
  that cannot be read is named on the same artifact-visible stream instead of
  vanishing. scripts/analyze-shellcheck.sh gets the same treatment for
  *.sh / *.bats / *.bash (both its git diff and its git ls-files run with
  core.quotePath=false, so a non-ASCII name — tracked or untracked — is not
  silently dropped), keyed on untrackedness so a tracked file
  whose only hunks are pure deletions keeps its empty range, with an
  untracked (whole-file scope): header line in its artifact. A --files run
  never widens beyond its pathspecs, .gitignored files stay out, and a
  tracked-only tree's output is byte-identical to before.

• run-suite.sh runs bats test files in parallel for a project that opts in
  (#593). New [project.<name>] key suite_jobs (integer ≥ 1, default 1 =
  serial) passes --jobs N --no-parallelize-within-files to bats; tests inside
  one file still run in order, matching the file-granularity safety audit in
  Issue-593's analysis dir. DEVAGENT_SUITE_JOBS overrides it for one run and
  is scrubbed from both suite children (and unset per test by
  tests/lib/hermetic-env.bash), so a test that invokes the runner cannot
  inherit it. Scrubbed alongside it: BATS_NUMBER_OF_PARALLEL_JOBS and
  BATS_NO_PARALLELIZE_ACROSS_FILES, which bats reads from the environment, so
  an inherited value would make the new bats_jobs: line false in either
  direction; and BATS_PARALLEL_BINARY_NAME, which would otherwise let bats
  reach for a binary the probe below never checked. The value is validated
  before either suite runs; N > 1 first checks for GNU parallel by banner (bats
  1.10.0's own probe is miswired — an absent binary surfaced inside the TAP
  stream as a "truncated" suite) and dies naming the package and the
  suite_jobs = 1 seam. The artifact gains a
  trailing bats_jobs: <N> | (none) line. Carriers: README "Running the test
  suite", tests/README.md "Parallel execution", the config skeleton,
  docs-site/configuration.md, the spec's run-suite section, and the
  core-draft-mr skill's artifact-read note.

• The rederive prober compares the checkout to its baseline and resolves bare
  basenames (#590). scripts/rederive.sh now opens its artifact with a
  ## Checkout vs baseline line — behind N, ahead M against default_baseline
  after a best-effort fetch whose outcome (ok / skipped / FAILED / n/a)
  is stamped on the line. Pre-branch, behind > 0 is a ✗ … STALE CHECKOUT
  falsified premise whose remedy names the merge --ff-only to run (Issue-570
  ran draft through improve three commits behind its base with every file probe
  ✓); on the recorded issue branch (every revision re-runs draft) the same
  numbers print as an ℹ row — expected drift, not a premise. An unconfigured
  or unresolvable base prints an explicit ? behind-count undetermined line,
  never a silent 0; a faulting rev-parse/rev-list/ls-tree dies loud. A
  .devagent-baseline marker is acknowledged by existence in the heading and
  never parsed (only branch.sh reads it). A backticked bare basename or path
  suffix (SKILL.md, capture/capture.sh) is matched against the HEAD tree by
  exact suffix at a / boundary: one hit → ✓ tok → path (resolved …) and the
  resolved path feeds the since-log and the file:line drift check; several →
  a ~ … ambiguous advisory row; none → the same ✗ as before (Issue-570's
  artifact carried six ✗ rows, zero real). Plumbing: upstream_fetch sets
  UPSTREAM_FETCH_STATUS (skipped|ok|failed) and runs its fetch with
  GIT_TERMINAL_PROMPT=0, so git's own credential prompt fails instead of
  hanging an unattended step (ship and mergetoall inherit that; an ssh prompt or
  a black-holed remote is not covered — a fetch timeout is a recorded
  follow-up). The prober measures the issue's recorded worktree when
  use_worktree put the branch there (#571's active_tree_resolve), else
  source_dir. Carriers:
  commands/draft.md, the imPlan template's Preconditions note, core-scope
  question 7, and the commands/branch.md override rules describe the new rows.

• The checklist writers fail closed (#589). checklist_mark refuses a
  name-less mark of a step number that is absent from the ACTIVE ## Revision
  block but present in an older one — the write used to fall th...