e6457cc849
Both review lenses PASSED; this is the cleanup, not a rescue. Evidence:
docs/research/2026-08-06-ap156-review-closure.md (the review itself is
committed alongside it as the received artifact).
R1 — the load-bearing containment test could not fail. Its truth and flood
values were two hand-copies of the same expression over the same part set,
so the shortfall was algebraically identically zero for any DAT input. The
oracle is now PHYSICS-POLYGON VERTICES — a different DAT field from the
bounding sphere the builder emits, so the two sides can genuinely disagree.
Sabotage-verified three ways after full cleans: dropping the bounds centre
in production reddens it (428 Setups, worst 35.869 m on 0x0200129A, matching
an independent out-of-repo sweep exactly); dropping only the scale on the
centre reddens it (326); and corrupting the TEST's own bounds oracle reddens
it (467) where under the shipped oracle that same corruption was invisible
by algebra. Renamed accordingly. A6's stale "cap control" comment corrected:
that loop is the test's own uncapped re-implementation and cannot observe a
cap regression — the cap is covered in Core.
R2 — the population was understated. 172 is AP-152's DISPATCH population;
AP-156's is 530 BSP-bearing Setups, of which 525 have a flood sphere move
and 428 fail vertex containment before the fix (412 at a 1 cm tolerance —
the review's figure; the gap is 16 Setups between 1.4 mm and 10 mm, real
geometry). 0 fail after, at any tolerance down to zero. Corrected in the
AP-156 row, the section-3 header, the C5c handoff and two test docstrings.
Dated review artifacts are left as written — "170 of 172" was correct for
what they measured, and rewriting evidence to match a later measurement
loses provenance.
A1 — BoundsCenter = default reopened at the type what the commit closed at
the seam. Dropping the default alone would NOT have closed the review's own
scenario (a copied Cylinder call site would write Vector3.Zero explicitly
and stay green), so ShadowShape's constructor is now private and BSP shapes
are built only through ShadowShape.Bsp(..., FlatCollisionSphere localBounds),
which takes radius and centre as ONE value and scales them together. There
is no expression a caller can write that carries one and drops the other.
22 construction sites converted; the same sabotage now reddens 5 Core tests
where the review's sabotage A reached 4, because both BSP producers share
one scaling path.
A2 — #333 is real and bigger than filed, and its retail question is
answered. I disassembled CObjCell::find_obj_collisions @0x0052b750 from the
PDB-paired binary myself (check_exe_pdb.py MATCH) rather than inheriting the
claim: its only early-out is sphere_path.insert_type == INITIAL_PLACEMENT_
INSERT, then it calls FindObjCollisions on every unparented non-self shadow
object UNCONDITIONALLY. Retail has NO distance pre-filter, so acdream's
"+ movement + 2f" reach filter is an invention with no register row — filed
as AP-158, carrying the disassembly, the F_EPSILON = 0.0002 m contrast, and
the measured blast radius (118 of 477 unique installed physics-BSP GfxObjs
exceed its ~2.5 m budget, 46 exceed 5 m). Active AP rows 109 -> 110.
Recorded prominently in three places a reader will hit: TALL PROPS MAY SHOW
NO VISIBLE CHANGE UNTIL #333 LANDS, and a null result at the connected gate
is EXPECTED, not evidence against AP-156.
LOW items. R3: the comment claiming the cited evidence justified the whole
cap line is corrected, but int.MaxValue on the sorting-sphere branch stays —
capping at 1 would take Spheres[0], and retail's one sphere is
CSetup::sorting_sphere, a different DAT field; capping keeps the wrong field
AND flips the substitution under-inclusive (#98/#168 direction). AP-157
already owns it. R4: acdream scales the flood sphere where retail's
find_transit_cells never reads gfxobj_scale — added as a second residual on
AP-156. R5: retail's slack constant carried into AP-158 and #333. A3: the
per-call delegate allocation is back to a cached field, still derived from
the single bounds resolver. A5: noted; b52967de's message cannot be amended.
Gates: all 44 bin/obj deleted before every verdict-deciding build, each test
run gated on a verified "Build succeeded" in the same invocation. Release
build 0 errors / 21 pre-existing warnings. Complete suite 11,208 passed /
4 skipped / 0 failed — reconciles exactly with the e2b2d04c baseline; one
test renamed, none added, removed or skipped. Nothing conflated with the
known load-sensitive flakes #302 / #308 / #321.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
311 lines
16 KiB
Markdown
311 lines
16 KiB
Markdown
# C5c closeout + successor handoff — the placement cutover's automated half is done (2026-08-06)
|
||
|
||
**Read this before any C5c or post-campaign work.** It is the successor to
|
||
`2026-08-05-c4-closeout-handoff.md` (whose ⚠ BISECT HAZARD block still
|
||
applies and is repeated in §6).
|
||
|
||
Branch `claude/acdream-physics-divergence-5aa784`, 21 commits from `02578441`
|
||
to `7b3e2895`. **Nothing is pushed** — the branch does not exist on the remote,
|
||
and there are 388+ unpushed commits ahead of `origin/main`.
|
||
|
||
---
|
||
|
||
## 1. One-paragraph state
|
||
|
||
Every implementation item in the placement cutover campaign is now landed and
|
||
dual-reviewed. **C5b** (classify-before-merge), **#280** (portal destination
|
||
prefetch), **#276's remainder**, **AP-22** and **AD-10** all shipped with both
|
||
review lenses PASS. **#309** was accepted as a standing divergence by user
|
||
decision rather than fixed. What remains for C5c is **entirely connected/visual
|
||
work plus the ledger close** — none of it can be done without the user at the
|
||
client, and the probe strip cannot be done before it.
|
||
|
||
---
|
||
|
||
## 2. What landed, by slice
|
||
|
||
| Slice | Commits | Outcome |
|
||
|---|---|---|
|
||
| **C5b** — classify before merge (#275) | `735f0a72`, `ed806997`, `23aa62f2`, `ff100cf3`, `9ee9c1a1` | Retired **AP-131** and **AD-60**'s legacy half. Seven production lines; five commits, because the review found a headless regression the change itself introduced. |
|
||
| **#280** — portal destination prefetch | `3aab05b0`, `73cdb95c`, `bcb66ccd` | Reveal window now derives from the live streaming radii. **D-1**, an unrecoverable portal hang, was found by review and fixed. |
|
||
| **#276 remainder** | `408c8e8f`, `fafc0b65` | Settle now adopts the transition's resolved cell across an indoor seam. |
|
||
| **#316**, **#317** | `429775d4`, `1d2d4bb8` | Report-only investigations. #316 **cosmetic**; #317 **no retail basis**. |
|
||
| **#309** | `43cfdc4a` | **Accepted as a standing divergence** (user decision), not a planned fix. AP-136 is its permanent record. |
|
||
| **AP-22** | `bc4679cd`, `619de97a`, `ef976c6d` | Invented collision cylinder deleted in **all three** copies. Row retired. |
|
||
| **AD-10** | `fe6ee877`, `886333a2`, `fb454b74`, `2223ed17`, `7b3e2895` | **Retired by deletion** — its stated justification was false at HEAD. |
|
||
|
||
---
|
||
|
||
## 2.5 C5c's automated gate — PASS
|
||
|
||
Run at `7b3e2895` on the final binary, **after deleting all 44 `bin`/`obj`
|
||
directories** (see §7 rule 4 — this session had three stale-artifact
|
||
incidents, so an incremental result would not have been evidence):
|
||
|
||
```
|
||
dotnet test AcDream.slnx -c Release -m:1 # ACDREAM_PAK_PATH set
|
||
```
|
||
|
||
**11,196 passed / 4 skipped / 0 failed**, all nine projects:
|
||
|
||
| Project | Passed | Skipped |
|
||
|---|---|---|
|
||
| AcDream.App.Tests | 4,172 | 3 |
|
||
| AcDream.Core.Tests | 4,261 | 1 |
|
||
| AcDream.Runtime.Tests | 1,220 | 0 |
|
||
| AcDream.Core.Net.Tests | 764 | 0 |
|
||
| AcDream.UI.Abstractions.Tests | 546 | 0 |
|
||
| AcDream.Content.Tests | 125 | 0 |
|
||
| AcDream.Headless.Tests | 89 | 0 |
|
||
| AcDream.Bake.Tests | 15 | 0 |
|
||
| AcDream.Cli.Tests | 4 | 0 |
|
||
|
||
The 4 skips are the pre-existing set; no new skip was added anywhere in the
|
||
campaign. **None of the three known load-sensitive flakes (#302, #308, #321)
|
||
fired** in this run — but they are separately filed and must never be
|
||
conflated if one does.
|
||
|
||
Net movement across the campaign: 11,106 at `02578441` → **11,196**, +90.
|
||
|
||
---
|
||
|
||
## 2.6 Connected gate — #280 USER-PASSED 2026-08-06; the rest NOT RUN
|
||
|
||
**#280's reveal gate: PASS, user-accepted.** User's words: *"now portal space
|
||
takes longer but terrain is complete when I exit."* Both halves are the
|
||
criteria the gate specified — a measurably longer hold (the fix doing its job)
|
||
and a complete destination on reveal (the acceptance).
|
||
|
||
Probe evidence, `ACDREAM_PROBE_PARK=1`, retail UI, Release, `ACDREAM_STREAM_RADIUS`
|
||
unset (`c5c-gate-after2.log`):
|
||
|
||
- **Three Portal reveals plus a Login reveal, every one at `radius=12`.**
|
||
Pre-#280 this was a hardcoded `1`.
|
||
- Each portal hold raised the wait cue at ~5.0 s (`elapsedMs=5031 / 5000 /
|
||
5010`) before `materialized → world-visible → complete`.
|
||
- Generation 4 revealed cell `0x3032001C` — the same cell as generation 1's
|
||
login, i.e. a repeat visit to an already-seen landblock.
|
||
- The park path was exercised (95 `[park...]` lines).
|
||
|
||
**An accidental but genuine A/B.** An earlier run in the same session set
|
||
`ACDREAM_PROBE_REVEAL_RADIUS=1`, which is a radius VALUE, not an on/off flag —
|
||
so it forced the pre-fix window. The user observed the original defect
|
||
(landscape building in the background) with `radius=1` on every probe line, and
|
||
did not observe it with `radius=12`. That is the before/after pair the gate
|
||
asked for, obtained by accident. **Successors: this probe overrides the radius;
|
||
it does not merely enable logging.**
|
||
|
||
### NOT RUN — do not read this section as gate coverage
|
||
|
||
| Owed gate | Status |
|
||
|---|---|
|
||
| D-1's two reachability scenarios (double-recall to the same landblock with a walk between; mid-hold quality-preset drop) | **NOT RUN.** Generation 4's repeat visit is suggestive but is not the demote path D-1 needs. |
|
||
| AP-136's six-step park check | **NOT RUN as the six-step protocol.** The probe fired 95 times incidentally; that is not the check. |
|
||
| Route-7 thickening | **NOT RUN** — `ACDREAM_PROBE_REMOTE_TELEPORT` recorded **0** lines; no remote teleport occurred. |
|
||
| Two-client observation | **NOT RUN** |
|
||
| Canonical nine-stop soak | **NOT RUN** |
|
||
| Lifecycle/reconnect route on the final binary | **NOT RUN** |
|
||
| AD-65 / AD-66 local-player visual gate | **NOT RUN** |
|
||
|
||
The campaign ledger closes with these outstanding **by user direction**, not
|
||
because they were discharged. Anyone citing "C5c passed" must cite §2.6's table
|
||
alongside it.
|
||
|
||
---
|
||
|
||
## 3. WHAT C5C STILL OWES — all of it needs the user
|
||
|
||
Nothing below can be discharged without a live client. The automated half is
|
||
complete.
|
||
|
||
### 3.1 Connected gates, batched into one sitting
|
||
|
||
1. **#280's reveal gate.** The user's original repro was a **recall**, so the
|
||
route needs a **lifestone leg**, not only `/teleloc`, plus a first-login
|
||
stop. A/B with `ACDREAM_PROBE_REVEAL_RADIUS`, `ACDREAM_STREAM_RADIUS`
|
||
unset. The pre-fix run is *expected to show the defect*; the post-fix hold
|
||
must be measurably **longer**. If it isn't, the gate widened nothing.
|
||
2. **#280 D-1's two reachability scenarios, never reproduced live**: two
|
||
consecutive recalls to the **same** landblock with walking in between, and
|
||
a mid-hold quality-preset drop. Both are fixed and unit-covered; neither has
|
||
been seen on a running client.
|
||
3. **AP-136's six-step park check** — `ACDREAM_PROBE_PARK=1`. **This survives
|
||
#309's deferral**: it validates the SHIPPED rollback path
|
||
(`restorableOnCancel` in `SubmitPreparedPlacementCore`, the shared core
|
||
behind every production placement), not the deferred fix.
|
||
4. **Route-7 thickening** — its gate passed on one `cause=propagate` probe line;
|
||
the evidence is THIN.
|
||
5. **AD-65 / AD-66** need a local-player visual gate before any fix (see §5).
|
||
6. **C5c proper**: two-client observation, canonical nine-stop soak,
|
||
lifecycle/reconnect route on the final binary, and the user's visual matrix.
|
||
7. **AP-156 — the BSP flood-sphere placement fix (`b52967de`).**
|
||
**The criterion is CONTAINMENT, not size, and a null result on tall props is
|
||
EXPECTED.** Two things must be said to the user before this one is run, or
|
||
its outcome will be misread in both directions:
|
||
- *"A prop stopped blocking"* is **not** by itself a bug report. AP-156
|
||
shrinks the flood for 143 of the 172 AP-152 Setups on purpose, because
|
||
retail's BSP branch (`calc_cross_cells` @0x00515230, `0x0051528f jne`)
|
||
contributes **no primitive at all** — the old flood was larger only
|
||
because it included a sphere retail never reads. The question to ask of
|
||
any such observation is whether the object's own BSP geometry still
|
||
reaches the cell it stopped blocking from. If it does, that is a bug; if
|
||
it does not, that is retail.
|
||
- *"A tall prop STILL does not block"* is the expected symptom of
|
||
**AP-158 / #333**, not of this fix. AP-156 puts the geometry in the right
|
||
cell; acdream's own `maxReach` broadphase filter — which retail does not
|
||
have at all — then discards it one layer down, for 118 of the 477 unique
|
||
installed physics-BSP GfxObjs. **A null result on tall props is not
|
||
evidence against AP-156.** Run this gate after #333, or run it on
|
||
short/wide off-centre props where the offset is inside the ~2.5 m budget.
|
||
|
||
Population, for the user: the change touches **525 of the 530** BSP-bearing
|
||
Setups (not the 172 the commit body names — that is AP-152's dispatch
|
||
population). See the AP-156 register row.
|
||
|
||
### 3.2 Then, and only then
|
||
|
||
**The probe strip — DEFERRED 2026-08-06, deliberately, and the ledger closes
|
||
without it.**
|
||
|
||
Six flags — `ACDREAM_PROBE_REMOTE_LANDING`, `ACDREAM_PROBE_REMOTE_SLIDE`,
|
||
`ACDREAM_PROBE_PARK`, `ACDREAM_PROBE_REMOTE_TELEPORT`,
|
||
`ACDREAM_PROBE_CHILD_CELL`, `ACDREAM_PROBE_LOCAL_TELEPORT` — plus
|
||
`ACDREAM_PROBE_REVEAL_RADIUS`.
|
||
|
||
**Not stripped, because §2.6's table says their gates were never run.** Only
|
||
#280's reveal gate was discharged. Route-7 thickening recorded zero
|
||
remote-teleport lines; AP-136's six-step park protocol was not performed; the
|
||
two-client observation, nine-stop soak and lifecycle/reconnect route did not
|
||
happen. Stripping the family now would delete exactly the instrumentation
|
||
those still-owed gates need, which is the failure this section was written to
|
||
prevent — so honouring it means *not* stripping, even though the campaign is
|
||
closing.
|
||
|
||
`ACDREAM_PROBE_REVEAL_RADIUS` is kept too, despite #280 being closed: **AP-149**
|
||
(outer ring accepts terrain-only) and **#326** (the missing Viewing Distance
|
||
option) are both open and both would want the same A/B harness. Retiring it now
|
||
would mean rebuilding it.
|
||
|
||
**Successor:** strip the family when the §2.6 gates are actually run, not on a
|
||
calendar. Note again that `ACDREAM_PROBE_REVEAL_RADIUS` is a radius **value**,
|
||
not a boolean — setting it to `1` silently reproduces the pre-#280 defect.
|
||
|
||
### 3.3 Then close the ledger
|
||
|
||
Register / roadmap / milestones / memory, and the campaign ledger close.
|
||
|
||
---
|
||
|
||
## 4. Issues filed this session
|
||
|
||
| # | Subject |
|
||
|---|---|
|
||
| **#321** | `DatSoundCacheTests` concurrent-decode dedup, full-suite load (third load-sensitive flake) |
|
||
| **#322** | Two callers compute the same pre-placement flags from the same inputs |
|
||
| **#323** | A far-snap store can silently stale a pending initial-create receipt |
|
||
| **#324** | Graphical and no-window hosts run parallel, non-shared inbound routes |
|
||
| **#325** | **Gate A's teleport test is narrower than retail's** — `==` where retail is "not older" |
|
||
| **#326** | acdream has no Viewing Distance option (retail's `Render.LandscapeDrawDistance`) |
|
||
| **#327** | No analogue of retail's DDD prefetch progress readout |
|
||
| **#328** | Camera far plane hardcoded 5000 f; retail's `zfar` is byte-verified 4000 |
|
||
| **#329** | Portal wait cue arms 5 s late; retail emits per tunnel rotation segment |
|
||
| **#330** | **Headless registers no live-entity collision at all** — a bot walks through every NPC |
|
||
| **#331** | **`ResolveWithTransition` refuses ALL uphill motion with a body supplied** |
|
||
| **#332** | Headless bots appear to have no remote dead-reckoning |
|
||
|
||
**#331 is the one to look at first.** Severity was raised from UNKNOWN once the
|
||
discriminator was found: it is the **`body:` parameter**, not the fixture. With
|
||
`body: null` the uphill sweep climbs; with a body it returns `ok=False` and zero
|
||
movement — under a call profile identical to the local player's
|
||
(`IsPlayer | EdgeSlide`, human two-sphere Setup), on ramps as shallow as **1.1°**.
|
||
A diagonal request keeps cross-slope X and zeroes only up-slope Y. Nothing in
|
||
the suite asserts uphill progress on a walkable slope, which is why it was
|
||
invisible — the test that found it passed **vacuously**.
|
||
|
||
---
|
||
|
||
## 5. Register movement
|
||
|
||
**Retired:** AP-1, AD-1 (C5a) · AP-145 (C5a) · **AP-131**, **AD-60**'s legacy
|
||
half (C5b) · **AP-22** · **AD-10**.
|
||
|
||
**Filed:** AP-147 (delta-stream cardinality) · **AP-148** (#325's Gate A
|
||
narrowing) · AP-149 (outer ring accepts terrain-only) · **AP-150** (wait-cue
|
||
delay is not retail's trigger) · AP-151 (gate stricter than retail on the
|
||
GPU-upload axis) · **AP-152** (live path emits primitives *and* BSP additively
|
||
where retail is exclusive; 172/5,935 Setups incl. BSP doors) · AD-64 (the
|
||
duplicated residency decision) · **AD-65**, **AD-66**.
|
||
|
||
**AD-65 deserves attention.** Its magnitude was filed at half the truth: the row
|
||
states `cos²θ` but quantified `1−cosθ`. Corrected to **25% short at 30°, 50% at
|
||
45°**, confirmed by measurement (#331's probe: 0.0735 m for a 0.1 m request at
|
||
30.96° = `cos²(30.96)`). **It is a live lead for #269's slope-slide residual** —
|
||
and note that project memory's #269 do-not-retry covers *friction and jump
|
||
chains*, which are byte-exonerated; `AdjustOffset` is a different function and
|
||
is **not** covered by it.
|
||
|
||
---
|
||
|
||
## 6. ⚠ BISECT HAZARD — carried forward
|
||
|
||
Commits **`735f0a72..23aa62f2`** contain a live headless defect: every remote
|
||
entity's `FullCellId` is frozen at its placement value for the whole session in
|
||
`AcDream.Headless`, and the local player loses one of AP-146's three
|
||
cell-refresh edges. Introduced by `735f0a72`, fixed at `ff100cf3`. Nothing
|
||
throws; no test in the range catches it.
|
||
|
||
---
|
||
|
||
## 7. Process findings — stated as rules, each paid for this session
|
||
|
||
1. **A blast-radius enumeration only reaches as far as the call graph its author
|
||
walked.** C5b's was performed over the graphical `OnPosition` path and missed
|
||
`AcDream.Headless` entirely — 11,000 green tests, one frozen host. Ask *which*
|
||
traversal, and what it structurally could not reach. Both hosts, every time.
|
||
2. **Two independent adversarial reviewers converging is near-proof; a lone
|
||
finding is a lead.** It happened three times this session (the headless hole,
|
||
the missing payload gate, D-1) and all three were real.
|
||
3. **Assume a test does not discriminate until sabotage proves it.** SEVEN green
|
||
tests covering nothing were found or avoided: C5b's conservation test, #276's
|
||
three settler tests, #280's tautological integration fixture, the atlas-tier
|
||
seam 4,170 tests missed, AD-10's only existing test (a dead method with a
|
||
hard-coded formula), and AD-10's contract-specified T1 sabotage which came
|
||
back green and was rejected rather than shipped.
|
||
4. **`bin`/`obj` can serve deleted code even under `--no-incremental` and
|
||
`-t:Rebuild`.** Three incidents. Delete all 44 directories before any
|
||
verdict-deciding result. A stale artifact does not look like an error — it
|
||
produces a plausible failure, or a plausible pass.
|
||
5. **Every contract in this campaign has been wrong at least once, and the
|
||
implementation is what catches it.** C5b's §3-D2 (a "dead" ternary that was
|
||
live), #280's §7 (three false items), AP-22's §7.1 (a literal that was a
|
||
three-way condition), AD-10's §7.1 (a non-discriminating sabotage). Brief
|
||
implementers to rebut, and treat a reasoned rebuttal as more valuable than a
|
||
compliant edit.
|
||
6. **Verify a cited address is the construct you claim.** AP-150 mis-cited
|
||
`0x004D7064` — a `PStringBase` constructor — as the `SendNotice` call,
|
||
*despite* being filed with a byte-level disassembly. Precision of method does
|
||
not prevent an error of line.
|
||
7. **Binary Ninja drops flag tests.** `if (-((eax_7 - eax_7)) == 0)` renders an
|
||
always-true where a real wrap-safe compare lives. Confirmed at Gate A
|
||
(0x00454054), `DoVectorUpdate` (0x004521F5), `HandlePlayerTeleport`
|
||
(0x00452186), and `FindObjCollisions`. Disassemble the PDB-paired binary
|
||
wherever a comparison or constant is load-bearing.
|
||
8. **Do not fan out subagents.** Two blowups: six agents spawning their own
|
||
children exhausted a session usage limit and killed four tasks mid-flight;
|
||
a later three spawned five more. Every brief must say **"do not spawn
|
||
subagents"** explicitly — none of the early ones did. One code-writer at a
|
||
time; the shared worktree tolerates no more.
|
||
9. **A shell failure inside a compound command can leave a commit claiming work
|
||
it did not do.** `ef976c6d` was needed because a heredoc invoked `python`
|
||
(absent here; `py` is the binary) while the commit still reported success.
|
||
|
||
---
|
||
|
||
## 8. Where to start
|
||
|
||
- **Post-campaign:** #331 first (§4), then AP-152 and #330 — both are real
|
||
collision divergences with user-visible consequences.
|
||
- **Domain entry points remain** `claude-memory/project_physics_collision_digest.md`
|
||
and `claude-memory/project_render_pipeline_digest.md`.
|
||
- **The campaign plan** (`docs/plans/2026-08-02-placement-cutover.md`) item 5 now
|
||
records AP-22 and AD-10 as retired; item 3 records #280 as done.
|