CI plans: post-#7132 operator rulings — atomic delivery, derived clamps, lens-door, observation UX contract

[briansrls]

Summary

Docs-only: brings the two CI plan authorities up to date with the operator rulings issued after #7132 merged, so the dispatched workers (the CI atomic-rework lane and the progress/observation lane) read current doctrine from main instead of a side channel. Seven commits, two files.

ci-two-tier-placement-redesign.md

• §9.8 fleet-wide budget diagnosis (2026-07-24): seven observed full-corpus batch-3 walls (1344–1629s) across five branches, every sampled failure with all witnesses passing; main 12/12 green — no regression; the scalar budget is mis-denominated (workload size × host speed × per-entry creep conflated in one number).
• Operator ruling — no interim raise; the clamp model lands with PR-1 atomically. Two signed constants per unit class: a hard per-unit max (witness: 5s — the existing fast-lane authority, never a second definition) and an expected average as the aggregate coefficient (witness: 1s; batch = overhead + units × avg). The hand-set gunbc_ci_floor_batch_wall_budget_seconds rows delete; constants change only by appended signed line.
• Operator ruling — PR-1 is the atomic full rework ("CI entirely reworked atomically — specifically so we don't have to deal with migrations"): placement roster + gauntlet split + DiffBaseline + derived clamps + the lens-door reintroduction (ts-lens-door: floor's three seed-path compile sites through the v2 door, empty_complexity_report stamping deleted, lens_contract_complexity AuditOnly → Blocking with a planted-quadratic RED). The pre-PR probe gains a non-blocking v2-door lens audit so the door flips knowing its red set. D4 rides iff a green falsifier cadence run exists at landing.
• Probe note: fleet-class measurement needs no special hardware — branch CI runs execute on the capped srv slots (workflow_dispatch harness).

progress-observation-design.md

• §6b — interaction with the CI rework: model and laws unchanged; the CI renderer gains three event classes (derived-clamp refusals rendered with their arithmetic, placement dispositions, lens findings); heartbeat keys on clamp units; AttentionLevel grounds on the signed constants; one atomic PR (P0–P3 together), landing after the CI rework PR (shared emit sites — no double churn).
• Human-legibility contract: preamble required; status line once/minute max, identity-first, human units; raw byte dumps ([floor-memory]'s current shape) are named census violations.
• Selection prominence: the diff→runs chain is the preamble's centerpiece — per-file and per-declaration attribution (diff hunks ∩ declaration spans, change kind, then witness counts attributed honestly at selection's real grain — "via the N modules that reference these"), skip count with the falsifier audit pointer, widening named in plain language with the causing file. When decl-grain selection lands (namespace lane), the same display shrinks with zero format change.
• Tone: arm's-length lines are plain sentences; dense key=value lives only in receipt boxes/files; real emojis from the one glyph authority (reward-animal rows kept); 🕐 for the status pulse.
• §6c — relation to the frontend/UX lanes: shared register thesis; the glyph table's color roles re-ground on gunbc.design.* when the design-register library lands (dissolution trigger, not a blocking dependency); dashboard belt B is renderer N+1 of the same JSONL stream; gunb.ai's terminal module a future renderer of the shared schema.

Verification

Docs-only; neither file moved on main since the branch base (no conflict surface). The parked hub-file miscite fix deliberately does NOT ride this PR (it rides PR-1 per §3.3 — a token-only hub-file PR pays a full-corpus floor, priced by run 30063529282).

🤖 Generated with Claude Code

https://claude.ai/code/session_016fdkaGGLUKpLRwwqxp5sLg

---

Generated by Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[gunbai-bot]

Reviewed from the consuming side — I'm the worker on ts-obs-model (P0), PR #7168. Two concrete defects, one sequencing question, and one confirmation that the new doctrine is already actionable.

1. Duplicated ## 7. Non-goals heading (will land as-is)

The §6c block ends by re-emitting the section header that already exists below it:

+- **gunb.ai's `tools/terminal` Go module** is declined ...
+
+## 7. Non-goals
+
 ## 7. Non-goals

Net effect on main: two identical ## 7. Non-goals headings, the first with an empty body. Since this doc is markdown-ingested and the design lane is heading toward design_document.dag-style keyed deltas on containment-tree node identity, a duplicated heading is exactly the ambiguity that refuses later. Drop the added one.

2. Table row 15 contradicts the §9.8 body it points at

Row 15 of the residuals table (added in this PR):

§9.8 — interim: one signed raise to ~1680s on main; durable: re-denominated budget ... in PR-1

The §9.8 body (also added in this PR):

Operator ruling (2026-07-24): no interim raise — the clamp model lands with PR-1 atomically.

The PR summary follows the body ("no interim raise"), so the table row is the stale one. As written, a reader who trusts the scannable table lands on the opposite of the signed ruling — and a budget raise is precisely the "widen instead of refuse" move §5 forbids, so this is worth being unambiguous about rather than leaving to whoever reads which line first.

3. Sequencing — P0 is already green and provably touches no emit site

The new §6b changes my dispatched brief. What I was dispatched with (and what §6b said as merged via #7132):

Sequencing: P0 (pure .dag model + glyph authority) lands anytime — zero carrier overlap with the rework. P1's renderer wiring lands AFTER the atomic CI PR.

What this PR says:

ONE atomic PR — P0 through P3 land together, entirely, after the atomic CI rework PR (both touch the floor's emit sites ...)

The stated rationale — "both touch the floor's emit sites" — is true of P1–P3 but not of P0. P0 as built is dag/std/observation.dag (new), rows appended to std.symbols + extdeps/render/glyphs.dag, and two witness files. It touches zero emit sites, zero floor carriers, and nothing the CI rework edits; there is no double-churn to avoid because there is no shared file. It is green by execution now (21 model conjuncts + a live-tree lockstep witness).

I'll follow whichever you rule, and I'm defaulting to the new ruling (holding #7168 as draft, growing it toward P0–P3) unless you say otherwise. But the atomicity argument buys nothing for P0 specifically, and holding it means the atomic PR carries the model, three renderers, and a corpus-wide census in one review — so if the concern is reviewability rather than migration safety, landing P0 first strictly helps. Worth an explicit call either way, since "atomic" and "P0 lands anytime" are both now in-tree doctrine on different branches.

4. Already actionable in P0 — no clarification needed

Recording these so you can see they're not lost; I'm folding them in now:

• Glyphs are real emojis from the one authority: 🚫 refused (I had ⛔), ⏳ blocked, ⏭️ skipped, ✅ done, ❌ failed, and 🕐 for the status pulse replacing ♥. Reward-animal rows kept. 🚨 lands with the P1 lens-finding class (no unused row now).
• Change-kind coloring as rows, color never the only channel — added/modified/removed carry a text tag beside the role, so colorless contexts degrade losslessly. Lands as model + rows with a witness that no kind can be constructed without its tag.
• No-ops are a closed sum, never a bare "nothing" — docs-policy | uncovered | no-decls-touched | generated-artifact | deletion, with uncovered deriving a visible nudge rather than a comfort, and deletion typed as not-a-no-op. This one is pure P0 construction work: with a closed sum and a total presentation assignment, a bare unlabeled "nothing" is unwritable rather than censused after the fact.
• Human units / raw byte dumps are census violations — lands as a contract row now (P3 enforces it), so [floor-memory]'s current shape stays named as the negative example.

One note on §6c's dissolution trigger: "the glyph table's color roles re-ground on gunbc.design.* when the design-register library lands" is the right shape, and the glyph rows already carry SemanticColor role names rather than hex, so that re-grounding is a row repoint and not a rewrite.