The tier budget gets a reader, and ten branch sections agree with the code again

[OpenCnid]

Two commits. The first gives the density-trellis's declared tier budget a reader and a floor. The second brings ten stale branch sections back into agreement with the code, and names who reads a note.

The budget had no reader

tier_budget_words: 90 occurred exactly once in this repository — its own declaration in docs/density-chain/index.json. Nothing read it, so the map drifted to tiers of roughly 800 words while every gate stayed green. A declared constant with no reader is a comment.

--budget now reads it and checks three clauses: every tier inside its band, a floor as well as a ceiling, and no class ending longer than it starts. The floor is the clause a ceiling-only rule cannot have, and it is the load-bearing one — a tier under budget was never compressed, and compression is the mechanism the format rests on (docs/density-chain/README.md; Adams et al. 2023, arXiv:2309.04269 §2).

--verify reads the field unconditionally and gates on it only when tier_budget_enforced is true, which is false today. The branches as shipped breached, and a gate that reddens CI on day one is a gate someone switches off — the same reasoning that keeps staleness out of --verify. Flipping the flag is an operator decision, not a side effect of this change.

The counting convention lives in extractTiers rather than in each caller because three separate readers counting one tier by eye returned 799, 803 and 844 words for the same bytes.

What the counter showed, which was not what was expected

Of the 49 breaches on first measurement, 31 were tiers under the floor and 14 over the ceiling. Two thirds of the drift was tiers that had never been compressed at all, not tiers that had grown. The one 799-word tier was the minority case.

Ten branch sections

Ten of thirteen classes were stale. One read-only sub-agent per class returned a densified replacement against that class's own declared paths, every claim carrying a commit, a path and line, or a command's output. Each return was verified by re-running those commands before it was applied.

Beyond currency, the pass found:

• Three commit hashes offered as receipts resolve to nothing. C5 diagnosed the cause: a pre-squash hash never survives into the history a later reader searches, and this repository squashes. The convention needs a fix, not the three hashes.
• C12's own status ledger contradicted the cross-section table on whether a production launch path exists. The code settles it for the table; the ledger carried two false sentences. The remaining gap is narrower and now locatable — install_scaffold and control still raise, so no guest supervisor has run inside a microVM.
• The map understated the system about as often as it overstated it. C4 found a flag described as having no consumer read by two live paths; C3 found a subsystem described as test-only driven by five operator commands.

Breaches fell 49 → 14, all remaining in C1, C2 and C13, which were never stale and are not in scope here.

Verification run

Every CI step, run locally on this branch at aa24d03:

check:repo-surface           PASS
wiki:check -- --verify       PASS
python:check                 PASS
test:textedit                PASS
npm test                     PASS — 1437 tests
npm run build                PASS

wiki:check --negative-control exits 3 with all 32 planted conditions detected, up from 25; the seven new ones cover the budget clauses.

Rule 19(c) — both new gates were watched failing before being trusted:

• Floor clause removed → negative-control BROKEN: the gate did not exhibit an_under_floor_tier_is_caught, exit 1.
• tier_budget_enforced set true → --verify FAIL, 49 issues, exit 1.

Both restored; the counts above are the restored state.