feat: a node that was never authored is never advised to freeze, and prose that claims output must prove it

[TinDang97]

The affordance-truth milestone, closed 7/7 and driven through ADD itself.

1 — A node that was never authored is never advised to freeze

direction.md states the design plainly — "There is no author verb — you fill those sections by editing that file directly" — and no affordance in the engine knew it. BEAT_NEXT["direction"] mapped the whole beat to add freeze {slug}, so from the moment add new Task wrote a file of placeholders, five surfaces recommended a verb freeze is structurally guaranteed to refuse.

The cost was never the wasted call. A scaffold and a finished contract read identically from status, so a node could sit unauthored for days inside a milestone that read as in progress — observed live in another repo, and true of this bundle's own v3-final-collateral since 2026-08-11.

The derived beat gains a scaffold rung between created and frozen, sorting first in todo under its own header:

$ add todo
1 open task(s):
scaffold:
  · fresh-thing → author fresh-thing's RULES, ASSUMPTIONS and CHECKS, then add freeze fresh-thing

It is computed from the same predicates the refusals call — gives_unauthored, placeholders_in — never a copy, because two notions of "authored" is precisely how advice and refusal came to disagree. status, todo, new and the CARD scaffold all resolve through one map, and status's own stamp test, a third reading of the beat, is gone.

On Milestones the bad advice was not refused — it SUCCEEDED. placeholders_in reads RULES · ASSUMPTIONS · CHECKS and a Milestone body carries none of those three, so it returned [] for every milestone and ADD's one human approval was stampable against a node stating no goal, no scope and no exit criterion. Measured, not inferred:

$ add new Milestone probe-ms && add freeze probe-ms
freeze recorded at authority `process`      # with goal: <one line> untouched

_milestone_stubs closes it, deliberately narrow: goal · why · EXIT, the three the milestone lifecycle already reads, since milestone_done refuses on why: and on the EXIT tally. A guard reaching SCOPE and GROUND too would refuse real milestones whose ground is thin, and a guard everyone learns to widen past is worse than a narrow one that holds. It refuses this repo's own v3-final-collateral — the intended one-time authoring cost.

card_drift now compares the DERIVED beat, not the raw status: field. This repairs a falsified assumption recorded by a replan on 2026-08-17: freeze does not move status:, so a freshly frozen node advertised next: add freeze <slug> — the approval it had just passed — while todo and status derived build, and doctor reported it CLEAN. render_card goes through _next_verb for the same reason; reading the map raw put a third answer on the CARD, missing that a sealed-but-unbriefed task owes add brief first.

2 — A claim that the engine prints something is proven by running it

promised-capability-guard closed this class for the READMEs, and its own why: named the gap it left: those guards check nouns the engine exposes, never capabilities the prose promises. loop.md is where that bit. It told the reader add status shows goal not met (m/n exit criteria), and pointed at "the plan-vs-state line in add status". status prints neither. The first string did not even live in add.py any more — it had been reworded to milestone_goal_unmet inside milestone_done's refusal — so an existence anchor of the kind the README guard uses would have resolved it and passed.

Both sentences sat in the loop's Gather step, and the first was named as THE cue that starts the loop, so an agent following the skill waited for a signal the engine never sends.

The guard registers every skill-tree line naming a backticked add <verb> and a rendering, and proves each by driving that command against a bundle put into the state the claim describes. The anchor is captured stdout. An unregistered claim fails by name, quoting the sentence with its file and line. Both false claims are reworded to truth in all three live skill trees; the new Gather cue is add milestone-done refusing with milestone_goal_unmet (m/n exit criteria), and planned-but-unscaffolded tasks are found at the scaffold beat this PR ships. Both replacements are themselves registered and driven.

add.py is untouched by that task, asserted by a check that diffs it — building the missing status surface is a real improvement and a separate task. A guard that ships having made its own claims true has never once refused anything.

Verification

The gate refused both tasks' first PASS, and was right both times. Two distinct ID traps, worth knowing about:

• a parametrized check binds nothing — pytest reports it as test_x[param] and the gate binds bare ids, so a green check sat there with its rule unbound;
• a module name binds nothing either (test_tree_parity is a file, not a test).

Beyond those, five probe: clauses had no check at all. Eight checks added across the two tasks, contracts refrozen, then gated.

• Red first: 12 of 21 checks failed against the pre-change engine for task 1. The green ones were deliberate counter-guards — a frozen node's affordances unchanged, the quick one-call lane still closing, the verb count, the status: vocabulary, frontier ordering.
• Dogfooded on a scratch bundle and on this repo's own.
• 818 passed / 7 skipped across both roots. Four engine twins synced, both MD5 pins re-aimed, py3.10 syntax floor checked with ast.feature_version.

Reviewer notes

• A frozen Reject from another task shaped the design. build-orient froze R:T2SCAN — "building a report must not read a single body" — and status needed the body. That Reject is not this task's to weaken, so _is_scaffold is one function with two tiers: T0 always, T2 only when the caller already holds the body. Disclosed residual: a node with an authored gives: but still-template RULES reads authored to status alone. todo and freeze both catch it, and it is recorded in the docstring.
• Two exit criteria were amended rather than ticked, each carrying its amendment note in the milestone body: criterion 1 said "every surface", which no surface bound by R:T2SCAN can satisfy; criterion 4 described the wider guard scope that was deliberately narrowed. Neither was ticked as written.
• The milestone closes checked by Tin Dang (unattended) — the ticks came from a process, not a human at a terminal, and the attribution shipped in feat: ceremony proportional to kind and size, and add check for the tally the gate reads #206 records that rather than letting a claimed name stand alone.
• test_new_scaffold.py's pinned affordance is re-aimed, not deleted; two incumbent value pins in test_todo.py and test_doctor.py likewise. The test_doctor fixture was replacing a string the body no longer carried — a silent no-op its own assertion caught.