chore(harness): adopt copier template v0.7.0

[egparedes]

Stacked on #11 — targets ao/ulv-3/template-v0.6.0, so this diff is the
v0.6.0 → v0.7.0 delta only. Merge #11 first.

Why this stacks on #11 rather than restarting from main

v0.7.0 requires the docs/ → development/ relocation #11 performed: it is
an upstream breaking change, and the new _skip_if_exists list protects
development/*.md paths only. Starting from main would mean redoing 77 files
of relocation to reach the same place. Exactly one commit of #11 is superseded
(ea24433, the hook repair), and that is a reconcile, not a reason to discard
the rest — see "Rejected hunks" below.

What v0.7.0 is

Two upstream ADRs:

• ADR 0013 — hook payload parsing. The Claude Code hooks had an
  undocumented hard jq dependency that degraded differently and silently at
  each of three call sites. New .agents/hooks/hook-input.sh parses with jq,
  falls back to python3 (probed by running it — stock macOS ships a stub
  that passes command -v), and returns distinct exit codes (3 = no parser,
  4 = bad payload) so each hook picks a posture: PreToolUse denies with the
  install remedy, Stop and PostToolUse skip with a visible non-zero exit.
• ADR 0012 — simplification wave. 24 questions → 13; the four role playbook
  skills merge into their subagent files; the SPIKE-/EXPLORER-/
  PLAN-REVISION markers collapse into one HANDBACK(kind)/RESULT(kind)
  convention; the example work unit, the verify skill, the three .agents/
  sub-READMEs, and the cursor/mcp modules stop being generated.

Eleven answers drop out of .copier-answers.yml; nothing re-prompts.

What survived untouched

.mcp.json, the Playwright block in .opencode/opencode.jsonc, all five real
work units, and scripts/fmt-file.sh — the last because v0.7.0 added
scripts/verify.sh and scripts/fmt-file.sh to _skip_if_exists, so the
working formatter from #11 is now protected rather than at risk.

Rejected hunks and how each was resolved

copier update --conflict rej produced five. Two were resolved in favour of
upstream outright:

Reject	Resolution
.claude/settings.json	Upstream wins. #11's own jq handling exited 0 on a jq-less host, letting every Bash command through unchecked; upstream denies with a remedy. Strictly better.
architect.md, product-owner.md	Dropped. The round caps #11 added by hand are now upstream text.

Three local hunks are re-applied on top:

• block-destructive.sh keeps this repo's matcher. Upstream's single
  substring grep denies any command that merely quotes a pattern — enough
  to block an rg search of the file itself, and it blocked two legitimate
  tool calls during the v0.6.0 adoption. The header now records the divergence
  and says to keep rejecting that hunk; behaviour stays pinned by 23 tests.
  This is the one deliberate, recurring cost in this PR: one .rej per
  template release on a 60-line file.
• PreToolUse stays anchored to ^Bash$. Unanchored it also matches
  BashOutput/KillShell, whose payloads carry no .tool_input.command — and
  upstream's new fail-closed empty-command branch would deny them. Verified:
  a BashOutput-shaped payload returns rc=2 through the unanchored wiring.
• The OpenCode formatter keeps bash, not sh. fmt-file.sh uses
  set -o pipefail and [[ ]]; under dash it aborts on line 1 and formats
  nothing, silently, because a formatter's exit code is not surfaced.

AGENTS.md also keeps its concrete license and squash-merge lines, which
v0.7.0 replaces with a _Fill in: marker and strategy-generic prose.

Preserved-doc drift

_skip_if_exists is matched at any depth, so README.md covers every README
in the tree. That left .agents/README.md pointing at three sub-READMEs and
four playbook skills the update had just deleted, and harness-usage.md still
teaching SPIKE-REQUEST: and the verify skill. All four affected files were
unmodified renders of the previous version, so they are replaced with the
v0.7.0 renders and the local content re-applied — the Browser automation (MCP)
section returns to harness-usage.md, and tool-bootstrap.md's two
_Fill in: markers are answered from pyproject.toml and the CI matrix.

development/style.md is deliberately not synced (see register row
2026-07-template-v0.7.0.3).

Verification

• make verify green: 376 tests, lint, format-check, docs build.
• tests/test_harness_config.py: 33 → 45 tests. The hook-wiring guards moved
  from asserting on inline jq to asserting the hook-input.sh contract and
  each hook's exit-3 posture; new TestHookInput covers the reader, including
  that the python3 fallback is byte-identical to jq — a backend that
  disagrees makes the guard machine-dependent.
• End-to-end through the real PreToolUse command string, with and without
  jq on PATH: benign allowed, destructive denied, quoted mention allowed.

Follow-up worth considering (not in this PR)

Three deny mechanisms still overlap and only one is the improved matcher:
.claude/settings.json's permissions.deny prefix rules and
.opencode/opencode.jsonc's substring globs remain the old style, so they
over-block relative to the hook. Either loosen them to match or upstream the
matcher to the template — the latter would also end the recurring reject.

🤖 Generated with Claude Code