v0.8.3 — three cost lessons from the second real run

Three cost lessons the second real ClickHouse run measured. All three were guidance the
docs got wrong or never gave, and each one cost real money or a rejected-but-correct
diff on the run that found it.

Changed

• resume guidance no longer implies resume is cheap. An executor's session is
  re-sent in full every turn, so each round pays the whole accumulated prefix again
  (measured across three attempts of one task: 7.69M → 9.18M → 9.35M input tokens; the
  third attempt — eight lines in 59 seconds — cost more input than the original
  1181-line implementation). The guidance now says: put every finding into one resume,
  and make trivial mechanical edits yourself.
• max_changed_lines now has authoring guidance. The cap had rejected correct work
  twice, both times sized to the implementation while the contract demanded three tiers
  of tests. Size it to the real diff shape: tests and deletions count too.
• The stage gate now costs the build step. A warm build directory does not make
  verification cheap: adding one new source file re-triggered CMake's
  CONFIGURE_DEPENDS glob and invalidated 9,891 object files against a four-hour CI
  budget. The gate now checks the project's own build budget, flags changes that add
  files rather than only editing them, and — when the promised verification cannot
  run — says "this was never compiled" in those words.

v0.8.2 — lessons from the first real end-to-end run

Two report/tidiness defects from the first real end-to-end run — both found by running
a real ClickHouse task through the whole loop rather than by reading the code.

Fixed

• router usage no longer suggests routing the orchestrator row. That row is the
  main model by definition — the one row that cannot be routed — and suggesting it next
  to otherwise-correct hints cost the report its credibility. Suggestions now consider
  dispatch rows only.
• land no longer leaves an empty .router/worktrees/<id>/ shell behind. The
  directory grew one empty entry per task that had ever run and read like a list of
  live runs. It is now dropped with rmdir, which refuses a non-empty directory, so a
  second run of the same task is never touched.

v0.8.1 — ship the state-guard fix

Fixed

• Shipped the state-guard fix (#51). The guard now allows what the workflow asks
  the orchestrator to write (TASK_CONTEXT.md, gate.yaml); without the version bump
  the fix stayed on main while every installed plugin kept refusing those writes.
• dist/router.js rebuilt — the bundle takes its version from package.json at build
  time.

v0.8.0 — the version installed users update to

main had carried the v2 round (#37–#48) while the plugin manifest still said 0.7.0 —
and that number decides whether an installed plugin sees an update at all.

Added

• The v2 workflow round: plan namespacing under .router/plans/<plan_id>/ with
  spec.lock, the plans verb, the optional TASK_CONTEXT.md mechanism (off by
  default, measured), queue-gate discovery and plan revision, and the widened executor
  gate grant that survives a resume.
• docs/ added to the packaged files, since the README points at docs/workflow.md.

Changed

• Plugin/command descriptions rewritten to state the actual division of labor: the CLI
  owns mechanism (isolated worktrees, supervision, concurrent dispatch, the
  exclusive-lock verification queue, environment-free gates); the main session owns
  every judgment, including the pass/fail verdict.