feat(vibe): Test named mechanisms and decision provenance in vibe sessions

[shafty023]

Purpose

#298 taught cl-execute to weigh a mechanism a ticket names against what already exists, to trace recorded decisions to a human, and to keep human-decision findings out of the planner's hands. Vibe sessions had none of that, so a writer could still build whatever mechanism a ticket named, read a Decisions section as settled, and a reviewer finding that needed a human could be quietly absorbed. This ports the same intent to vibe without adding a plan review, a lint script or a build-loop step, and without asking the vibe person any engineering question.

What changed

• Guardrails reviewer (handoff only) gains a Mechanisms check: for each new poller, interval, background lane, socket, cross-process channel, route, table, cursor, or write into another tool's files, name the existing mechanism that could do the job and why it was not used, state what its cost scales with at 10x and 100x plus the per-device idle cost, and never argue from fleet size. Idle-scaled cost when an activity-scaled design meets the outcome, or a write into another tool's data with no cited human decision, is blocking. A ticket-named mechanism gets the same test, and the plan's Cost Scaling table is checked against the diff.
• New Engineering decision: <question>; built: <...>; evidence: <...> item. Handoff section 6 and quality-loop.md: a finding that a mechanism, a one-way door, or an unratified or conflicting decision needs a human decision is never the writer's to resolve, downgrade or absorb; it goes on the live ticket's Engineering checklist, never to the person. Ticket template and ticket worker place it.
• Change worker: only the outcome binds; a named mechanism is a proposal. When it is dominated, the writer builds the better option and logs the swap as an Engineering decision (never NEEDS_PERSON). Plans that add periodic or cross-surface work carry a Cost Scaling table using cl-execute's planning-and-gates.md columns and rules.
• Requirements worker tags named mechanisms mechanism (proposal) instead of acceptance criteria.
• quality-loop.md defines settled: a decision traces to a human statement (who, where, which question). Unratified or conflicting ones go through the existing product necessity gate or become an Engineering decision; never resolved by recency. Every "never re-ask a settled decision" now points at that definition.
• closedloop-graph.md gains a route for channels that already connect two surfaces.
• Frozen rows describe user-visible behavior, never mechanism.
• vibe 1.0.47 to 1.0.48, CHANGELOG entry.

Commands run

node --test <every plugins/**/*.test.{mjs,cjs,js} file, as CI does>   # 392 pass
uv run pytest plugins/ -q                                            # 2316 passed, 3 skipped
uv run ruff check . && uv run pyright                                 # clean
cd tools/plugin-manifests && npm ci && npm test                       # 6 passed

New test_vibe_mechanism_provenance_contract.py pins the rules above; each test fails when its pinned phrase is removed. Three existing tests pin the Engineering checklist row in the ticket template's ownership table; that row now also lists engineering decisions, and the assertions were updated to the new row (no rule removed). The lean loop, same-family and no-plan-review contract tests pass unchanged.

Second commit: tools/python/vibe_contract_text.py now holds the flat, section, paragraph and agent helpers that 22 vibe contract tests each kept a copy of. Where copies differed, the shared section stops at the same or a higher heading level and matches its heading at line start (exact line preferred), paragraph splits list items, and agent returns raw text (three files now call flat(agent(...))). Every call site whose output changed was compared against its old helper, and only positive assertions read those sites. pytest plugins/vibe: 252 passed; pytest plugins/: 2316 passed, 3 skipped.