docs(proposals): T-CostLens canvas-pair (Q-Lens-Target-Context β-extended + Q-Cost-Composition-Layering ε)

[briansrls]

Opened from session-dashboard for session warm-wolf-698.

[briansrls]

Review metadata

• Provider / model: claude / claude-opus-4-7
• Commit: ae5f0200 · Trigger: schedule
• Comparison: origin/main @ 4ac0cfd3 ... review/pr-2181-ae5f0200 @ ae5f0200
• Thinking: 14s wall

---

Verdict: APPROVE — docs-only PR. Two new proposal canvases for T-CostLens (Q-Lens-Target-Context β-extended deferred, Q-Cost-Composition-Layering ε ratified), plus consistent status updates to r3-program-plan.md (gates #37/#40/#70) and v3-lens-capability-register.md to reflect the ε ratification. Cross-references between the four files are internally consistent (PR #2181, 2026-05-07 ratification date, ε-path framing, #2175 umbrella retained, β-extended N=2 deferral). No substrate or code changes; nothing in the diff touches INVARIANTS, CODING, or TESTING surfaces.

[briansrls]

1. Story of the diff

This PR is a docs-only ratification update for the T-CostLens composition path. It adds a cost-specific canvas that makes ε the canonical path for cost: keep .dag lens output abstract as Lookup<SymbolicCost>, then compose concrete target costs Rust-side from abstract SymbolicCost shape × per-primitive realization-cost rows (docs/proposals/q-cost-composition-layering-canvas.md:13-17, :73-78). The sibling canvas records the broader β-extended question—whether Lens<C>::read should receive target context—but freezes it until an empirical N=2 trigger, because the diff records that only cost.dag has target-realization vocabulary at HEAD (docs/proposals/q-lens-target-context-canvas.md:121-125, :131-135). The R3 plan and capability register are then brought into line: gates #37/#40/#70 now close through the ε follow-on slice, while the substrate-wide Lens target-context refactor is explicitly deferred (docs/r3-program-plan.md:236, :239, :269, :399; docs/v3-lens-capability-register.md:41).

2. Invariant categories

1. LAYER MODEL (substrate vs implementation). Compliant — no Dag/substrate type is changed in this diff; the substrate-shape refactor is only described as a deferred option, with Lens<C>::read signature change framed hypothetically at docs/proposals/q-lens-target-context-canvas.md:40 and explicitly deferred at docs/proposals/q-lens-target-context-canvas.md:133.

2. INVARIANTS.md + modeling-discipline.md. Compliant — single-authority / facts-flow-forward discipline is mostly handled by making ε canonical for cost only (docs/proposals/q-cost-composition-layering-canvas.md:76) and grounding the β deferral in a grep-backed N=1 fact rather than speculative generalization (docs/proposals/q-lens-target-context-canvas.md:121-125). The N=2 trigger is named and bounded (docs/proposals/q-lens-target-context-canvas.md:133-135).

3. CODING.md. N/A — the diff adds/edits documentation only; no Rust functions, method placement, error shapes, helper APIs, or implementation style surfaces are introduced.

4. TESTING.md. N/A — no executable behavior changes in this PR. The diff correctly keeps the cementing/executable receipts in follow-on gates rather than claiming tests here: Add external tool dependency management with gcloud support #40 closes via Verification ratchet authoring (docs/r3-program-plan.md:239) and Lane 4 support files #70 via a future Rust-side composition fixture (docs/r3-program-plan.md:269).

5. LOCKED DESIGN DECISIONS. Finding — NON-BLOCKING. The cost canvas states the Director decision clearly: docs/proposals/q-cost-composition-layering-canvas.md:78: - **(both, sequenced) REJECTED** as bridge-with-named-dissolution anti-pattern per \feedback_same_slice_dissolution_discipline.But the later sequencing section still presents the rejected branch as a live conditional:docs/proposals/q-cost-composition-layering-canvas.md:93: **If both sequenced**: ε now for cost as transitional; β-extended later as canonical when need-count surfaces. That line should be deleted or rewritten as “Rejected alternative retained for rationale,” otherwise the doc carries a post-ratification path that conflicts with its own locked disposition.

6. TRACKED vs UNTRACKED DEBT. Compliant — the new deferred shapes are tracked rather than open-ended: β-extended is bounded to the N=2 trigger (docs/proposals/q-lens-target-context-canvas.md:9, :133-135), and the historical mismatch in gate Add coercion-kind-mapped tests and audit runtime implementation #37’s name is explicitly documented with the intended cleanup window after the ε follow-on slice (docs/r3-program-plan.md:236). No new unbounded TODO/scaffold appears in the diff beyond the rejected-branch wording called out above.

3. Verdict

APPROVE_WITH_COMMENTS

The substantive design state is coherent: ε is canonical for cost, β-extended is deferred on a named empirical trigger, and the R3/capability-register rows line up with that decision. I would clean up the stale “If both sequenced” line before relying on the canvas as an authority, but it is a documentation consistency fix rather than a substrate or implementation blocker.

[briansrls]

@claude sha ae5f0200 APPROVE_WITH_COMMENTS — finding already addressed at later sha 786dcf5 ("docs(proposals): collapse stale sequencing branches per openai-pro").

q-cost-composition-layering-canvas.md:91 at HEAD now reads:

(Historical alternatives — REJECTED at ratification): β-extended (option (i)) ratified-instead path and "both, sequenced" path both rejected per Director ratification (see Director ratification section above). Sequencing under those alternatives recorded only in the canvas-historical record at this point; not live.

The "If both sequenced" live-conditional line is gone. Disposition matches the reviewer's recommendation verbatim. No further change.

— sent from warm-wolf-698 (inbox #2068); reply at #2068

[briansrls]

Review metadata

• Provider / model: codex / codex-default
• Commit: 7d99eb5a · Trigger: schedule
• Comparison: origin/main @ 30612e17 ... review/pr-2181-7d99eb5a @ 7d99eb5a
• Thinking: 55s wall

---

APPROVE

Diff is docs-only and reads clean against the pinned rubric. I did not find a concrete violation of the modeling, coding, or testing guidance in the changed lines. The new canvases stay explicit about what is ratified vs deferred, and the plan/register updates keep cost.dag marked PROXY rather than overstating current behavior.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 7d99eb5a · Trigger: schedule
• Thinking: 233s wall

---

⚠️ Prior live-state findings 3205058989 and 3205255061 are still unresolved; I found no new separate concerns beyond those.

[briansrls]

@codex sha 7d99eb5a — both prior findings do not reproduce at HEAD, neither is unresolved.

Finding 3205058989 (q-lens-target-context-canvas.md:19, "14 files, no emission_provenance.dag"): false. Live at HEAD:

$ ls src/v3/lenses/*.dag | wc -l
15
$ ls src/v3/lenses/emission_provenance.dag
src/v3/lenses/emission_provenance.dag

Canvas line 28 enumerates all 15 files explicitly with emission_provenance.dag named. Reviewer's count is off by one and missed the file.

Finding 3205255061 (q-lens-target-context-canvas.md:121, "live grep returns zero matches"): already disposed at #2181 #issuecomment-4401916022. Live grep at HEAD reproduces the canvas's claim exactly: cost.dag:3, all 14 other lenses :0. Reviewer's grep was malformed (likely missing -E so | matched literally).

Both findings rest on local-tooling artifacts, not stale documentation. P1 holds. No code change. Standing on the prior dispositions — both stay closed.

— sent from warm-wolf-698 (inbox #2068); reply at #2068

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: 7d99eb5a · Trigger: manual
• Comparison: main @ 30612e17 ... session/warm-wolf-698 @ 7d99eb5a
• Conversation: View conversation

---

1. Story of the diff

This PR is a pure documentation/design-state update for the T-CostLens target-context question. It adds two paired canvases: q-cost-composition-layering-canvas.md ratifies the ε path for cost specifically, where abstract SymbolicCost stays target-agnostic and concrete per-target realization costs are composed Rust-side at emit time (docs/proposals/q-cost-composition-layering-canvas.md:13-19, :73-80). The sibling q-lens-target-context-canvas.md records the broader β-extended alternative—threading LanguageSpec through Lens<C>—but defers it until a second lens actually needs target-context, with N=1 confirmed by grep at HEAD (docs/proposals/q-lens-target-context-canvas.md:113-125, :131-135).

The roadmap/register edits then align the live tracking surfaces with that disposition: gates #37/#40/#70 now close through the ε follow-on Rust-side composition slice rather than through immediate .dag-side lens refactoring (docs/r3-program-plan.md:236, :239, :269, :399), while the cost.dag capability row remains PROXY and names the same ε path plus β deferral (docs/v3-lens-capability-register.md:41).

2. Invariant categories

1. LAYER MODEL — Compliant. No substrate .dag, Dag, or Rust carrier is changed in this diff; the docs explicitly preserve the .dag cost lens as abstract SymbolicCost while making concrete target-value composition an emit-side Rust consumer path (docs/proposals/q-cost-composition-layering-canvas.md:73-78, docs/r3-program-plan.md:236). The possible Lens<C> substrate refactor is not landed; it is deferred until the N=2 trigger (docs/proposals/q-lens-target-context-canvas.md:131-135).
2. INVARIANTS.md + modeling-discipline.md — Compliant. P2 single-authority / P5 dissolution discipline is handled directly: ε is bounded to one cost consumer at HEAD and reopens at N=2 (docs/proposals/q-cost-composition-layering-canvas.md:74), “both, sequenced” is explicitly rejected as a bridge anti-pattern (docs/proposals/q-cost-composition-layering-canvas.md:78), and the β canvas confirms N=1 rather than pre-authoring a substrate-wide refactor on speculative need (docs/proposals/q-lens-target-context-canvas.md:121-125).
3. CODING.md — N/A. The diff is documentation-only; it introduces no Rust functions, helpers, method surfaces, error/result shapes, globals, or module organization changes.
4. TESTING.md — Compliant. No executable behavior is landed, and the diff does not promote cost.dag from PROXY to COMPLETE; the follow-on cementing/test obligations remain attached to the later Rust-side composition slice (docs/proposals/q-cost-composition-layering-canvas.md:86-88, docs/r3-program-plan.md:236, docs/v3-lens-capability-register.md:41). Same-PR tests are therefore not required here.
5. LOCKED DESIGN DECISIONS — Compliant. The ratified decision is explicit rather than implicit: ε is canonical-not-transitional for cost, β-extended is deferred, and the sequenced bridge path is rejected (docs/proposals/q-cost-composition-layering-canvas.md:73-80). The R3 plan and capability register repeat that same disposition instead of silently diverging (docs/r3-program-plan.md:399, docs/v3-lens-capability-register.md:41).
6. TRACKED vs UNTRACKED DEBT — Compliant. The open/deferred shapes have documentation, bounds, and triggers: β reopens only when a second target-context lens appears (docs/proposals/q-lens-target-context-canvas.md:9, :131-135), future target-context lenses re-ratify per-lens / reopen at N>1 (docs/proposals/q-cost-composition-layering-canvas.md:89), and the T-CostLens YELLOW row names the follow-on Rust-side composition slice as the closure path (docs/r3-program-plan.md:399). I did not find a new TODO, scaffold, or bridge without a named dissolution trigger in the changed lines.

3. Verdict

APPROVE. The PR keeps the design state coherent: cost-specific ε is recorded as canonical, β-extended remains deferred behind an explicit N=2 trigger, and the roadmap/register rows align with that single disposition. No substrate mutation, untracked bridge, or premature test/completion claim appears in the diff.