M0: expose segment attraction and scale dining time by courses

[HSpector1]

What this is

The two authorized M0 finishing corrections from the PR #7 adjudication (H2 legibility + M1 economics). Focused, in-scope; does not reopen DemandModel.Capture. Targets the PR #7 branch for a clean diff.

Do not merge. Review surface + human-gate decision. No human tests. No M1.

H-LEG-1 — attraction legibility

The menu-responsive demand model is now visible. SegmentAttraction (pool / consideration% / mix% / parties) is on ForecastSnapshot and ServiceResult:

• Pre-service: a "Who this plan attracts" table + one plain-language note (largest-remainder reconciled to the market).
• Post-service: an ATTRACTION funnel, so forecast vs actual arrival mix can be compared.
• New causal categories (disjoint-PState attribution, no double-count): weak-attraction (too few), wrong-mix (a composition opportunity cost — under-attracted the market's higher-spending crowd while the operation ran fine), purchase-conversion (arrived but didn't order), alongside seating/kitchen. Example: a value chassis in the enthusiast market now reads "wrong mix — under-attracted food enthusiast, the market's higher-spending crowd (~$204 missed), while filling the room with a lower-value crowd."

MED-ECON-1 — course-scaled seat duration

Flat EatMin=12 → shared Tuning.SeatEatMin(coursesX10) = clamp(6 + 6·courses, 12, 30) → 1/2/3-course = 12 / 18 / 24. The sim scales a party's eat window by DistinctCourseCount; the forecast scales seatCoverCap by the captured-mix expected courses (same primitive). Only eat time changes — AvgTicketTimeMin (kitchen) is untouched.

Economic effect (empirical): a premium 3-course service turns 2.96 covers/seat (~61-min dwell) vs a value 1-course 3.67 covers/seat (~33-min dwell) — premium ties up a table ~1.85× longer per seat-hour, so per-seat-minute leverage falls from ~4× toward ~2.5–3×. No over-correction: premium still wins enthusiast, value still wins lunch, mixed viable in social, KNOWN_RESIDUAL holds. Favorable side effect: best-generalist regret rose 6%→17% on base 271828182 (exactly the mechanism the adjudication predicted), not gate-chasing.

Technical

• 146 tests pass (Core 54 + Determinism 31 + Scenario 61, incl. 7 new ReadinessTests), 0 fail, stable across two runs; Release build clean.
• Determinism/accounting/RNG/forecast-immutability/scope preserved. Goldens deliberately re-baselined (duration changes seat occupancy; attraction legibility does not affect the checksum) — see D-032.

Gate recommendation

Verdict: Pass-with-notes
Action:  Continue   (request a focused independent re-review of these two corrections)

Continue does not authorize human tests, M1, or merging.

Where to look

• reports/m0/M0-HUMAN-GATE-READINESS-REPORT.md — §A–J incl. exact reviewer experiments (§J)
• docs/design/M0-HUMAN-GATE-READINESS-CONTRACT.md — the pre-code design lock
• docs/DECISION-LOG.md — D-031 / D-032
• Reproduce: --report (attraction table + causal categories), dotnet test

🤖 Generated with Claude Code

[HSpector1]

Final Focused Independent Re-Review — M0 Human-Gate Readiness (PR #9)

Independent technical readiness review of the two authorized corrections (H-LEG-1 attraction
legibility, MED-ECON-1 course-scaled seat duration) before the five-player M0 human gate.
Read-only. No branch modified, no commit, no merge, no human tests run.

A. Review identity

Repository	HSpector1/Restaurant
PR	#9 — M0: expose segment attraction and scale dining time by courses
Base	e5baa54 (PR #7 branch rewrite/m0-menu-responsive-demand)
Head reviewed	c360ebcc72ce1e0a59394e4368977feb29eca1ca (all 7 agents confirmed this exact SHA)
Date	2026-07-30
Environment	macOS (Darwin arm64), .NET SDK 8.0.423, Release config; fresh clone at /tmp/rev9
Method	Lead inline verification + 6 focused read-only sub-agents (A forecast-attraction, B funnel/causality, C course/seat timing, D forecast-timing, E strategy/economics, F docs/scope/tests), each on its own held-out seeds, then 1 adversarial cross-check of the one Blocker.
Verification loops used	0 of the 2 optional bounded loops (§30) — evidence was unambiguous.
Commands	dotnet build/test -c Release (x2); Harness --report/--probe/--attr; the interactive CLI; independent scratch drivers referencing the built RestaurantSim.Core.dll (public API) on held-out seed bases 424242000 / 511–558M / 522M / 533M / 544M / 555M.

Baseline confirmed: the diff (git diff e5baa54..HEAD) is 17 files / +662 −124 — 6 production .cs (Tuning, Simulator, Forecast, Result, TextReport + Harness/Program.cs diagnostics), 3 test .cs (AttributionTests, ScenarioTests modified; ReadinessTests new), and docs/reports. No binaries/assets, no M1 systems, Demand.cs/PriceModel.cs untouched, PR #7 base and main unchanged. Contract M0-HUMAN-GATE-READINESS-CONTRACT.md was committed before the code (commit 771e2b8 precedes the feat commits).

---

B. Executive verdict

Technical verdict:        Pass-with-notes
Action:                   Continue  (authorize the five-player M0 human gate)
Human-test authorization: Authorized with stated limitations
PR #9 merge:              Do not merge yet (final human-gate candidate; hold until the gate)
Foundation merge:         Do not merge yet (only after the human gate)
Main merge:               Do not merge yet (only after the human gate; M1 remains separate)
M1 status:                UNAUTHORIZED

Both authorized corrections are implemented, in scope, deterministic, and independently verified. No Blocker or High defect survives scrutiny. The single Blocker raised by a sub-agent (attraction table "attracted > pool") was refuted on cross-check and re-adjudicated by the lead to a Medium legibility note: it is internally consistent with the simulator, matches the authorizing contract's exact formula, and does not teach a false strategic rule. The remaining items are legibility/wording edge cases and one economic-rationale disclosure — none makes attraction invisible or corrupts the funnel the gate relies on. This maps to §31 "Authorized with stated limitations."

Load-bearing claims reproduced independently (held-out seeds):

Claim	Verdict	Evidence
SeatEatMin 1/2/3-course = 12/18/24, floor 12, ceil 30, integer-only, no name/premium branch	Confirmed	Tuning.cs:35-36; direct calls (0/5/-5→12; 40/100/1000→30; 14→14). Int128 math, no float in the timing path.
Course scaling actually lengthens table occupancy in the authoritative sim	Confirmed	Seat-bound social-dinner, held-out seeds: 3-course serves fewer parties & loses more to capacity than mains-only, every seed (e.g. served 55→47, lostCap 20→40).
Only the eat window changes; AvgTicketTimeMin (kitchen) untouched	Confirmed	Ticket time recorded at delivery (Simulator.cs:204) before Eating; ReadinessTests locks it.
Funnel disjointness gen == served+lostCap+lostWait+lostMenu	Confirmed	0 violations across 56,700+ held-out runs (Agent B) + 1,500 (lead). Four terminal PStates partition every party; tail sweep at Simulator.cs:267.
Forecast attraction table reconciles (pool→ExpectedArrivals, attracted→attempts) exactly	Confirmed	15 scenario×strategy + 13 synthetic edge cases; sums exact in all.
Attraction notes derived from real values, no market/segment/strategy-name branch	Confirmed	Forecast.cs:213-231 branches on ConsiderationBp/CompositionBp/PoolParties; names only in display interpolation.
Causal categories disjoint, ranked by max not sum; premium never mislabeled wrong-mix	Confirmed	DiagnoseBottleneck (Simulator.cs:656-664); premium-in-enthusiast wrong-mix = 0/200 (lead) & 0/40 (test); broad+overload → kitchen, not weak-attraction.
H-LEG-1 code is economically inert (attraction visibility does not move the checksum)	Confirmed	Agent E rebuilt pre-fix core + patched only the EatUntil line → reproduces post-fix byte-for-byte.
No dominant strategy; generalist is a fragile compromise, not safe near-best everywhere	Confirmed	2/3 distinct market winners; best-generalist worst-market regret 14/18/22% on held-out searchbases (all >10%).
Docs honest: no-merge / no-human-test / no-M1 stated; cross-OS determinism marked UNVERIFIED	Confirmed	Contract, report, decision-log D-031, DETERMINISM-CONTRACT.md.
146 tests pass ×2, Release, 0 warnings; modified tests re-baselined/renamed, not weakened	Confirmed	Core 54 + Determinism 31 + Scenario 61; goldens changed for MED-ECON-1 (documented D-032); weak-demand→weak-attraction is a cause-string rename.

---

C. Attraction legibility (H-LEG-1)

Forecast table (TextReport.Forecast): renders segment | pool | appeal% | mix% | attracted with a total row. Pool sums exactly to ExpectedArrivals and attracted sums exactly to attempts (= round(ExpectedArrivals × conversion)), via a remainder-to-last-segment allocation. The mix column correctly has no printed total (so a 100.01% rounding case never renders a visible contradiction). The forecast is fully immutable/deterministic (byte-identical across 20 recomputes).

Actual funnel (autopsy): a DEMAND FUNNEL block (attempted parties / served parties+covers / lost-to-seating / walked-waiting / no-acceptable-dish, correctly labeled parties vs covers) plus an ATTRACTION table (pool/appeal/mix/arrived) and a FORECAST vs ACTUAL line. A player can compare expected mix vs actual arrival mix directly (e.g. forecast 7/78/13 vs actual 6/78/14). Seeded variance is framed as "an analytic approximation… gaps are expected and informative," not as a forecast error.

Player questions (§26) — answerable from the CLI alone:

1. Which segments are in this market? → Yes (pool column, all 3 segments, pre- and post-service).
2. Which segments do I attract? → Yes (appeal% + mix% + attracted counts).
3. Which am I failing to attract? → Yes (low appeal/mix + the "barely registers with X" note).
4. Is my problem attraction or service? → Yes (causal summary separates weak-attraction/wrong-mix from seating/kitchen; funnel shows where losses fall).
5. Did the actual mix resemble the forecast? → Yes (forecast table vs autopsy ATTRACTION table + FORECAST vs ACTUAL).
6. What should I change next? → Mostly (primary causal recommendation + most/least-useful dish + attraction note), with edge-case roughness (A2/A3/A4 below).

Terminology issues: "appeal" for consideration and "mix" for composition read cleanly. The one confusing point is the pool vs attracted relationship (finding A1).

---

D. Causal attribution

The autopsy's causal summary distinguishes five disjoint channels and reports the single largest by measured lost contribution (ranked by Math.Max, never summed), with the busiest station as a strictly secondary "condition, not necessarily the primary loss." Rendered examples (Harness --report):

• weak-attraction — "weak attraction — only N parties considered a visit versus roughly 150 the market could yield."
• wrong-mix — "the plan under-attracted food enthusiast, the market's higher-spending crowd (~$204.12 of missed contribution), while filling the room with a lower-value crowd. The operation ran fine…" (business-consequence line + clean-operation qualifier).
• purchase-conversion — seated but nothing acceptable to order.
• seating / kitchen — the operational channels.

Disjoint-state audit (§11): every generated party ends in exactly one terminal PState (Paid / WalkedSeat / WalkedFood / NoOrder); demandGenerated == served + lostCap + lostWait + lostMenu held with 0 violations across 56,700+ held-out runs (Agent B) and 1,500 (lead). weak-attraction reads ungenerated parties; wrong-mix is a composition opportunity cost (not a party bucket); so no contribution is double-counted. The earlier "utilization flip" defect is not recreated — in tiny-room wrong-mix cases the missed-opportunity cents genuinely exceed the small realized seat loss.

Controlled fixtures (lead, 200 held-out seeds each):

Plan	Market	PrimaryLossCause distribution
healthy value (52 seats)	lunch	kitchen 181 / seating 18 / wrong-mix 1
healthy value (60 seats)	lunch	kitchen 200
value-chassis	enthusiast	kitchen 150 / wrong-mix 42 / weak-attraction 8
premium	enthusiast	kitchen 197 / seating 3 / wrong-mix 0
repellent menu	social	weak-attraction 200
tiny room	lunch	seating 199 / kitchen 1
broad + grill overload	social	kitchen 200 (not mislabeled weak-attraction)

wrong-mix is reachable but never fires for a well-matched premium plan (0/200), and only 0.5% for a healthy value plan (not a systematic false positive). Because M0 is throughput-dominated (D-012), "kitchen" is the primary cause in most services; the demand-side causes are narrow-regime by design, but the attraction table is always shown, so legibility does not depend on which cause fires.

Priority (§15): strong-attraction + weak-op → operational primary; weak-attraction + excellent-op → attraction primary with no add-capacity advice; weak + weak → both stages reported and ordered. Verified.

---

E. Timing correction (MED-ECON-1)

• SeatEatMin(coursesX10) = clamp(6 + 6·courses, 12, 30): 1/2/3-course = 12/18/24 exactly; 0/invalid → 12; ≥40 → 30 ceiling (never reached by the sim, whose max is 3 courses). Integer-only, name-agnostic, no premium penalty.
• Course-count determination: DistinctCourseCount counts distinct courses among ordered dishes (deduped, size-independent). A party always has a main before any starter/dessert is attempted (OrderDishes gating), so distinct ∈ {1,2,3}; a dishless party becomes NoOrder and never reaches the Eating state. Failed/comped courses still count toward the eat window (economically reasonable: the diner occupies the seat while a comped course is served).
• Total seated duration: measured in seat-bound 24-seat rooms — value 1-course ≈ 32.5 min dwell, premium 3-course ≈ 65 min dwell; a 3-course meal ties up a table ~2× longer. Multi-course parties demonstrably serve fewer parties and lose more to capacity.
• Forecast reuse: the forecast's seatCoverCap uses the same SeatEatMin (dwell = ForecastNonEatDwellMin(20) + SeatEatMin(expectedCoursesX10)), from the captured-mix expected course count. It cannot be reproduced by the old flat 12 (3-course dwell 38 / cap 217 vs old 32 / 258); a 1-course mix is byte-identical to legacy. Higher course count → longer dwell → fewer seat turns, monotonically.
• Table release / service close (§19): deterministic; occupied-at-close parties flushed to lost-wait (Simulator.cs:267); accounting non-negative and conserved on every held-out run; identical checksum on repeat.

---

F. Economic effects & premium-leverage ruling

Per-seat/per-station decomposition (base 555000000, n=200), premium retains every qualitative property: check $58.59 vs value $20.07; ingredient/cover $14.92 vs $4.53; ticket 24.2 min vs 11.8; effective dwell 65 vs 33 min; station util 63-64% vs 43-46%; contribution/cover $34.28 vs $9.94. Premium still wins the enthusiast evening; value still wins lunch; mixed viable in social. No over-correction; no correction that is economically inert.

Covers/seat clarification (§22): the builder's 3.67 (value) / 2.96 (premium) are theoretical seat-turn figures (= ServiceMinutes ÷ dwell) measured in controlled 24-seat seat-bound rooms — a basis the builder's report explicitly discloses (reproduced independently: 3.69 / 2.77). They are not realized covers/seat, which for the named strategies in their native markets are ≈ 3.21 / 2.51. The gate summary should carry this label so the figure is not over-generalized.

Premium-leverage ruling (§23): Credible with tuning notes.
The correction mechanically does what it should: each premium multi-course cover now ties up its seat 2× longer, so its per-cover-per-eat-minute leverage falls (≈3.4× → ≈2.2×). However, Agent E's isolation experiment (rebuild pre-fix core, patch only the EatUntil line) proved the more decision-relevant, undisclosed result: at the whole-service level in the enthusiast market — where premium is demand/kitchen-bound, not seat-bound — the longer dwell reduced wait-walkouts (LostToWait 10.8 → 4.4), so premium's enthusiast contribution rose +35–38% ($2120 → ~$2920) across 5 held-out bases, and whole-service premium/value leverage rose 1.29× → 1.80×. So the "premium now pays for its seat time / leverage reduced 4×→3×" framing is metric-dependent: it holds per-cover but inverts at the service level in premium's home market. The intended cross-market effect did land where it matters — the premium-anchored generalist's edge shrank (regret 6% → 14–22%). This is a rationale/disclosure gap, not a broken correction: strategy choice remains valid. The owners should be told the net effect on premium in enthusiast is a modest increase, not a taming.

---

G. Strategy regression

• Winners: lunch = Focused Value; social = Premium Craft; enthusiast = Premium Craft → 2/3 distinct, no single dominant strategy (distribution.md, base 555000000, 200 seeds; p10s reproduced).
• No cross-market dominator: best-generalist worst-market regret 14% / 18% / 22% on held-out searchbases 556/557/558M, all in enthusiast-evening, all >10%. Higher than the builder's stated ~6–17% band, which strengthens the no-dominator conclusion.
• Generalist ruling (§25): PASS. The generalist is a fragile, seat-time-consuming compromise, not the safe near-best everywhere. MED-ECON-1 nudged it less dominant (favorable). The underlying demand-rewrite residual (D-028/D-029) is a pre-existing PR M0 menu-responsive demand rewrite (Conditional-Pass; owner decision on residual) #7 owner decision, out of PR M0: expose segment attraction and scale dining time by courses #9's scope; PR M0: expose segment attraction and scale dining time by courses #9 does not worsen it. Per the mandate, I do not fail on a single sub-10% regret — and here every measured regret is above 10% anyway.

---

H. Technical regression & scope

• Tests: 146 pass twice in Release (Core 54, Determinism 31, Scenario 61 incl. 7 new ReadinessTests), 0 fail / 0 skip / 0 warnings. Debug and Release both green.
• Test integrity: modified tests were re-baselined/renamed, not weakened — 3 golden checksums changed because course-scaled seat occupancy changes covers/revenue (documented D-032; H-LEG-1 does not affect the checksum), and the weak-demand→weak-attraction rename keeps the same fixture and assertion. New ReadinessTests are non-trivial (SeatEatMin 12/18/24, ticket-time isolation, exact attraction reconciliation, causal fixtures) and reproduce on independent held-out seeds.
• Determinism / accounting / RNG / forecast immutability / numeric guards: all hold; forecast has no seed and is byte-stable; no float in authoritative timing.
• Scope (§28): grep confirms no executable reputation / repeat-visit / customer-memory / reviews / marketing / reservations / bussing / Godot / pathfinding / saves / managers / delegation / campaign / audio / assets. Clean.
• Documentation (§27): contract, readiness report, and decision-log all state PR M0: expose segment attraction and scale dining time by courses #9 awaits independent re-review and forbid merge / human-tests / M1; cross-OS determinism is correctly documented as UNVERIFIED/OPEN; the shipped example-forecast-vs-actual.txt reproduces byte-for-byte (checksum 05E65C6E8AD6AF07). One stale code comment (F-DOC-1 below).

---

I. Findings

(Severity: Blocker/High/Medium/Low/Note · Status = whether the tested claim held)

A1 — Forecast "attracted" can exceed a segment's "pool"

• Severity: Medium · Status: Confirmed (phenomenon) / the "no segment exceeds its pool" reconciliation check is not satisfied
• Category: legibility / labeling
• Claim tested: mandate §8 reconciliation — "No segment exceeds its potential pool."
• Evidence: attracted[s] = pool[s] × considerRaw[s]/Bp where raw consideration may exceed 100% (Demand.cs:109-127; Forecast.cs:200-203). On the default preset+market (Balanced Competent / social-dinner) the CLI renders Social Dinner pool 57 / attracted 59 (forecast) and pool 57 / arrived 59 (autopsy, seed 424242123); Broad Menu shows 57 → 74. Columns still reconcile exactly to ExpectedArrivals (95) and attempts, and it is consistent between forecast and sim. Matches the authorizing contract §2a formula exactly. ReadinessTests never asserts Parties ≤ PoolParties.
• Why it matters: on a legibility-focused gate, "attracted 74 from a pool of 57" reads as an arithmetic contradiction and the forecast table (unlike the autopsy table) carries no explanatory caption. It does not teach a false strategic rule — the lesson ("this menu is a social-diner magnet") is correct — and the aggregate funnel the gate relies on is sound.
• Required action: add a one-line caption to the forecast attraction table (as the autopsy table has) clarifying that a strongly-attracted segment can draw a share above its neutral market baseline, and/or relabel "pool" as "market baseline"; add a Parties-vs-Pool intent test. Owner decision, low-risk.
• Gate effect: Does not block. Stated limitation. (Sub-agent A rated Blocker citing the mandate checklist; adversarial cross-check refuted to Low citing the authorizing contract; lead adjudicates Medium — the mandate check genuinely fails, but against an assumption the authorized model does not share.)

A2 — Over-attraction note is silent in the unstaffed-station case

• Severity: Medium · Status: Confirmed · Category: notes / coverage gap
• Evidence: the overload note guard is kitchenCoverCap > 0 && attractedCovers > cap·1.2 (Forecast.cs:228). When a menu lists a dish whose station has no cook, kitchenCoverCap == 0, so the guard is false: ~107 parties attracted, ExpectedCovers 0, contribution ≈ −$1802, yet the only note is "mostly value lunch." The single most severe over-attraction case emits no overload warning.
• Why it matters: this is exactly the situation the note exists to catch. Mitigant: ExpectedCovers 0 and a large negative contribution are still shown (a loud separate signal).
• Required action: also fire (with distinct "kitchen can serve ~nobody" wording) when kitchenCoverCap == 0 && attractedCovers > 0.
• Gate effect: does not block; recommend polishing before or during the gate.

A3 — Self-contradictory note for a menu with no main course

• Severity: Medium · Status: Confirmed · Category: notes / contradiction
• Evidence: with no mains, Capture returns consideration=floor for all segments and composition falls back to raw pool mix; the ignored-branch then prints "barely registers with social dinner: only 1% considered it, so they are just 60% of arrivals despite being 60% of this market" — calling the largest arrival share "barely registers," with a "despite" between two equal percentages.
• Required action: guard the degenerate all-at-floor / no-mains case (e.g. "this menu has no main course; almost no one will choose you"); avoid "just N% … despite N%" when equal.
• Gate effect: does not block; reachable by a player who forgets a main.

A4 — Note fixates on a micro-segment for premium plans · Low · Confirmed

• For premium/enthusiast plans the prose note calls out value-lunch (4% of that market) instead of the enthusiast pull. The table still answers "who's most attracted"; only the one-line note is sub-optimal. Reorder note priority so a tiny-pool segment does not preempt the dominant/top-pull note.

A5 — "largest-remainder" label is really "remainder-to-last" · Note · Confirmed

• Forecast.cs:200-201 rounds the first N−1 and gives the exact remainder to the last segment. Safe for the current 3-segment fixtures; a future tiny trailing segment could underflow. Relabel the comment or implement true largest-remainder; add a ≥0 guard.

B-2 — Loss-making seating verdict prints "~$0.00 of lost contribution" yet advises adding seats · Low · Confirmed

• When a service is loss-making, cpcFloored == 0 so seatLoss == 0, but the count-based fallback still picks "seating" and appends "more seats would recover the most." Internally inconsistent for a non-expert. Suppress the cents figure / recover-by-expanding advice when the service is loss-making.

B-3 — purchase-conversion is a thin residual channel · Note · Refined

• Under menu-responsive demand, most would-be menu-rejecters are converted to non-arrivals scored as weak-attraction, so purchase-conversion is rarely PRIMARY (never mislabeled as seating/staffing). A design-intent clarification for the owners, not a defect.

B-4 — wrong-mix cents is a modeled opportunity estimate · Note · Confirmed

• wrongMixLoss = underCovers × spendGap × 0.60 is an estimate compared against realized cpc losses; disjointness/max-not-sum still hold. The line already says "missed contribution," which is fair.

D-3 — Forecast ExpectedCovers can point the wrong direction vs the sim on seat/course changes · Low · Refuted (as a timing defect)

• Forecast cover-direction disagrees with the sim in 5/9 paired tests, but every case traces to the documented kitchen-only realization keying (Forecast.cs:81-85, from PR M0 menu-responsive demand rewrite (Conditional-Pass; owner decision on residual) #7 / D-021) plus the separate economic model — not the timing primitive, whose seatCap/dwell math is directionally correct in all cases. The forecast is explicitly a low-confidence analytic estimate; the decision-relevant contribution direction agrees in the seat-tradeoff cases. Recommend the gate note that ExpectedCovers is a kitchen-ceiling estimate, not a realized-cover predictor.

D-4 — Seating-risk note is a kitchen-overload warning, not a seat-shortage warning · Note · Refined

• Forecast.cs:228-229 fires with LostToWait (kitchen cannot feed seated parties), matching its wording, not with LostToCapacity (seat-queue overflow). Internally consistent; optionally clarify in docs.

E-1 — Premium-leverage rationale is metric-dependent and partly inverted · Medium · Refined

• See §F. The builder's "leverage reduced 4×→3×" holds per-cover but inverts at the whole-service level in the enthusiast market, where premium contribution rose ~35–38% (emergent wait-walkout reduction). Correction is mechanically sound and strategy choice is valid; the disclosure is incomplete. Ask the builder to restate the effect as observed (or show the whole-service basis for ~4×). Not blocking.

E-2 — 3.67/2.96 covers/seat are controlled seat-turns, not realized · Low · Refined

• Theoretical ServiceMinutes/dwell in 24-seat seat-bound rooms (disclosed by the builder); realized named-strategy covers/seat are ≈3.21/2.51. Label accordingly in the gate summary.

F-DOC-1 — Stale code comment: "~70% / 62%-converting" vs the actual 35% floor · Low · Confirmed

• Simulator.cs:564-566 explains the weak-attraction floor as "~70% of the market's potential … a 62%-converting menu," but Tuning.WeakDemandCaptureFloorBp = 3500 (35%), recalibrated in the demand rewrite (D-030). The comment is pre-existing (blame d60f9646, PR M0 menu-responsive demand rewrite (Conditional-Pass; owner decision on residual) #7 base) but sits on the exact block PR M0: expose segment attraction and scale dining time by courses #9 renamed and edited. No runtime effect (code uses the constant). Correct the comment during the re-review cycle.

Confirmations that support readiness (Note/Confirmed): B-1 (conservation, 56,700+ runs, 0 violations), B-5 (one-course byte-identity + determinism), C-1..C-5 (timing primitive, distinct-course logic, seat coupling, table release, exact-2× dwell), D-1/D-2/D-5 (shared primitive, seats-past-kitchen-wall flat, ticket-time isolation), E-3/E-4 (generalist regret >10%, determinism/test integrity), F-SCOPE-1 (clean scope), F-TEST-1 (no test weakening).

---

J. Human-test ruling

Authorized with stated limitations.

Both corrections work; the remaining imperfections (A1 label, A2/A3 note edge cases, A4 note priority, E-1 rationale) do not teach false strategic rules; the generalist is a healthy fragile strategy; players can understand and test positioning from the CLI; only minor tuning/wording notes remain. Recommended (not required) before the gate: a small notes-polish pass (A1 caption, A2 unstaffed-kitchen warning, A3 no-mains phrasing) would raise fidelity on the very feature the gate tests — owner's call, since the core table is legible and the degenerate cases already surface loud 0-cover / negative-contribution signals.

---

K. Final gate

Verdict: Pass-with-notes
Action:  Continue

Continue authorizes the five-player M0 human gate. It does not authorize M1 and does not authorize merging.

Howard and Aaron should run the five-player uncoached M0 human gate using the reviewed PR #9 build (c360ebc).

Keep PR #9, the rewrite/correction chain, and the foundation unmerged until the human gate result is in; M1 remains separately unauthorized.

---

Independent review complete. Read-only throughout: no branch modified, no commit, no merge, no human tests run. This review reproduced every load-bearing claim on held-out seeds; where it disagreed with a sub-agent (A1 severity, E-1 status) the lead re-adjudicated with cited evidence.