[SamGu-NRX]

This PR fixes review comments on ten already-merged PRs that nobody addressed before merging. It covers everything outside the iPhone app: the reconstruction worker, the verification harness, the packet proposals, four experiments, the integration viewer and the docs. Each original PR has its own section below, linked to its threads. The remaining items include contract decisions, an unsafe photos-only bound needing a product tradeoff, and a fixable replay defect reassigned to the iOS caretaker. They are listed at the end.

#20: recon: handoff to the server team

• Free space could rest on one view (r4114128918, r4115044065).
  • Wall and ground coverage need two camera positions at least 0.25 m apart. Facing and overhead clearance only needed the fused volume's weight, which proves just that some frame's ray reached the voxel.
  • compute() now hands free_space a SeenFromTwo view of the volume. Its observed() also requires two such positions, each with a ray that reaches the sample according to that frame's own depth. The clearance march batches 16 slabs at a time; wall-front geometry receives the raw volume while clearance samples retain the two-view check. Visibility projections are bounded and use cached packed partner masks instead of repeated dense camera-pair multiplication.
  • A new test gets no clearance from one camera, none from two cameras 0.1 m apart, and 2.5 m from two cameras 0.4 m apart.
  • Current combined recon pytest suite: 158 passed on 6c0639e.
  • Caretaker microbenchmark: 24 cells, 32 frames and 5 cm voxels, three runs; median 0.7403 s before vs 0.3146 s after, 2,976 vs 192 visibility queries, identical answers. This small synthetic run does not establish full-capture throughput.

#240's ZIP admission follow-up is included: footer selection agrees with Python's reader, and malformed comment footers or ZIP64 locator overrides are refused before metadata allocation. Tiny archive regressions preserve the cap and normal-comment extraction.

The handoff now lists two-view free-space coverage among the fixed behaviors. All 11 visibility tests pass; the overhead footprint and photos-only uncertainty decisions remain open. Handoff review.

Vet follow-ups protect the effective ZIP extraction directory and its ancestors/descendants before startup cleanup. Cached members are verified by name, size and streamed CRC and damaged caches are re-extracted under a lock. Overhead samples include both edges of the existing 0.56 m footprint, catching the missed 1.90 m eave; support for other battery depths still needs the server footprint contract. At df7b75d, twelve new recon cases failed before these fixes; the guarded recon suite passes 125 tests on #240 and 149 on #242.

The Sol breadth follow-up closes the occupied-slab bypass: obstructions always stop clearance marches, but a measured gap and observed band require two-view support at the hit. One camera and two nearby cameras now export no unsupported measurement; two separated cameras retain supported measurements. Both edges of the existing 0.56 m footprint remain covered. Four controls fail on aaacc0f before this repair. Reported corner allowances must also be finite and nonnegative before rule adoption, with zero allowed; seven new controls fail on aaacc0f. Guarded suites on 6c0639e pass 474 verification, 158 recon, 99 meter-closeup and 20 drift tests.

Facing, overhead and wall-band samples now measure heights from ground.height_at(x,z) at each point, including slopes away from the wall. An uphill eave reported as 2.50 m before the fix is conservatively reported at 1.40 m above its cell ground. Observation at every slab still requires two separated camera positions. Local ground review.

#13: verification: independent end-to-end checks (S4)

• Hardcoded headroom (r4114226410). The ETH3D facade scene builder reads headroom.min_ft from server/rules.yaml at --rules-ref (default origin/t3/server, as the e2e harness uses) and records the ref and value in build_report.json. case.json's rules_assumed now lists headroom, so the existing mismatch check catches a private override through the threshold the server reports. BATTERY_HEIGHT_FT is still a constant. No thread raised it, but it is the same kind of problem.
• Unbounded response reads (r4114226421).
  • post() now uses http.client. One deadline covers DNS, TCP connect, TLS handshake, headers and body, and every status is capped at 4 MB. A bounded worker wait returns on deadline and closes sockets registered before TCP/TLS. OS DNS cannot be cancelled; a late result is checked before it can allocate or send. The cap's only evidence is that the largest result seen so far was 16 KB.
  • Tests run a local server that floods or trickles headers and body, on both 200 and 500.
  • Unlike urllib, http.client ignores http_proxy. Tests cover reduced remaining TCP/TLS budgets, blocked DNS/TCP/TLS cleanup and a localhost TLS peer that never completes its handshake.
• make sim-app passed on a shortcut (r4114865922).
  • The run must now log findMeter, meterCloseUp, wallWalk, markFeatures, uploading and result in that order. Each gapRequest must be followed later by an uploading. The names are ScanPhase's.
  • Tests accept the recorded runs at a39d0a5, f14947e and c820cf1. They reject an immediate result, a result with no upload, an unanswered gap request, and an early result hidden by a later complete flow.
  • Limit: a recapture the app never shows logs no state, so this check can't catch it.
• A new passing start slipped past the monotonicity check (r4114865928).
  • A start that the original sweep never evaluated now takes the outcomes of the original's neighbouring runs on that wall. This is sound because the solver evaluates every point where an outcome can change, plus the midpoints between them (server/solver.py:9-10).
  • A new pass is flagged only when neither neighbour passed, or when the wall had no run. The more-error check gets the same rule.
  • Regressions cover a pass in a fail gap, a pass on an unswept wall and a pass past the last run, and each fails on alpha. A boundary that legitimately moved still passes.
• Accessibility audit treated different screens as one (r4114865931, r4115043215). The screen key is now every element in tree order, by type, identifier and label. Unlabelled groups are skipped, and values and frames are left out. It typechecks against the iOS simulator SDK, but I didn't build or run the UI test.

python -m pytest -q in verification/: 554 passed on the combined branch. Ruff lint and formatting pass for verification and recon. Facade regressions retain clipped tail probes and the active headroom in the same observed entry.

This branch also contains #240's shared root Makefile dispatcher repair: verification is included, with standalone, missing-folder and failure-propagation cases in tests/makefile.sh. That helper is the cross-folder change required to make the verification suite run from the root.

The merged #240 deep-review repairs also require actual execution of all five hostile probes for full S2-8 credit. Generated and committed geometric cases bind their expectations to battery width and the applicable corner allowance; altered rules skip those expectations and do not establish acceptance. The shared root verifier target now runs locked dependency sync before child tests.

Codex found that a server-reported sweep such as [0, 1e9] could allocate billions of starts. Sweep admission now rejects spans outside the scanned chain before sampling, and the sampler validates eagerly and yields lazily with a 10,000-sample cap including the endpoint. Oversized runs fail with a specific message rather than get truncated. Regressions cover both outcomes, both scan ends, a distant singleton, tiny rule steps and the exact cap boundary. Sweep sampling review.

The rule-assumption follow-up compares case assumptions with reported spot width and rule-keyed thresholds. Geometry adopts a validated reported width only for a different policy fingerprint; a matching public fingerprint retains the loaded rules and checks reported dimensions against them. When the result omits a value that differs from the public rule, the report says "assumed public rules; server may override" and marks the record assumptions unverified. Rule-dependent checks are withheld, rule-independent failures remain failures, and JSON, Markdown and the CLI expose the gap without granting acceptance. An otherwise clean inconclusive run exits 2. All 392 verification tests pass, covering matching overrides, real reported mismatches, absent values, report/scoreboard gates and the existing hostile-span bound. No private backend files were read. Rule assumption review.

One runtime meter counts the geometry and comparison work that actually runs, including unsure checks, footprint lookups, ground grids, coverage and corner checks. Identical sweep records are deduplicated, and identical geometry checks are memoized. Structural admission rejects nonfinite numbers, overflowing span lengths, invalid dimensions and unknown wall references before geometry. The ground sampler is lazy; ordering comparisons use an interval index. A failed admission or invariant stops property comparisons, and any check exception becomes a named report problem. Exceeding the runtime limit produces verification work limit exceeded (N samples). Verification ZIP admission still counts directories before metadata allocation, responses remain bounded at 4 MiB, and required simulator states must occur in order.

The runtime limit is now 7,834,000 units, twenty times the remeasured valid peak of 391,700 on a saved ETH3D facade response. After caching, 35 valid public-fixture judgments including property variants peak at 31,649 units. Twenty additional measurements cover an 80-foot surveyed wall written as twenty walls, each subdivided 1, 4, 12, 13 or 20 times, original and mirrored, with full-judge resends. Those peak at 88,999 units. Another 468 distinct saved responses are included in the measurement corpus. These are operational work counts, not physical clearance accuracy.

The recorded public-solver response for an 80-foot wall has 683 candidates and passes with 29,760 actual visits and no ground-grid calls. Both missing-evidence branches use 130 visits for one unsure check or 10,000 duplicates. The old 2d959c7 prediction rejects the valid response and invokes geometry 10,000 times in either branch. The remeasured valid peak is 391,700 units.

Reported dimensions must be finite, positive and consistent with reported spans and footprint edges before rule adoption. A large reported width with no geometry to sample incurs only the work actually performed. Adopted dimensions and property variants charge the same judge meter. More error may add a failing run on a wall the original response never evaluated; a newly passing run remains a violation. Regressions cover invalid dimensions, span and footprint agreement, newly evaluated walls and bounded remaining response reads.

Current guarded suites pass: 594 verification, 168 recon, 99 meter-closeup and 20 drift-anatomy tests. All ran through run-guarded.py --min-free-gib 8 after disk met the 10 GiB startup threshold. Six sloped-ground controls and both reported-rule controls fail on 6c0639e before this repair. Ruff format checks pass in verification, recon and all touched experiment directories. The nine held formatting files preserve their Python syntax trees.

Unknown public-rule assumptions still withhold rule-dependent checks and case expectations, but identical-input, mirror and coverage comparisons continue. A nondeterministic response therefore fails verification even when the result omits its width. Rule-independent checks review.

#22: packet: point to the server's capture packet 0.4 and propose 0.5 additions

• Ground height in raw world coordinates (r4114080314). Proposal 2 now carries meterAnchor.groundOffset along the anchor's +y, so the ground moves with the meter when ARKit corrects the anchor. The app's local packet already records it this way (ground_y_m, PacketWriter.swift:412).

#59: experiments: signals lab round 1, which extra phone signals sharpen clearances

• A3 measured along travel, not along the wall's normal (r4113839737, r4115917036).
  • The axis across the walk is the wall's normal only while the walk runs straight along the wall. A new table restricts A3 to windows whose path stays within 0.25 m or 0.5 m of its start-to-end line. Both widths are untuned.
  • On those windows, along-travel error is 0.98 and 0.95 of the p90 on MARViN at 20 ft. This is a straight-path check, labeled throughout the code, table, README and METHODS. Without wall annotations it cannot identify wall-parallel travel or establish wall-plane benefit. The numerical results are unchanged; 20 metric/report tests pass.
  • Rerun against the evals harness at 190ed33. New strict tests are in tests/test_metrics.py.
• A4's anchor moves (r4113839728, r4115917030). Not a bug. Each window is one session that starts at the meter, and the second phone stays at the window's start. Sliding the start samples many meter placements. METHODS, the README and the generated report now say so.
• M3 picked the narrowest photo pair (r4115044359).
  • run.py --pair=widest reruns M3 with the widest qualifying pair and writes separate results.
  • On electro with modern_assumed poses, the 3-D p90 is 158.7 in for the widest pair against 154.6 for the narrowest, and 41% of points match either way. With exact poses the wide pair is worse, 156.4 in against 112.5, because fewer of its matches land within 2 px. P4 fails under both, so the narrow pairing isn't what makes M3 look bad.
  • Wide pairs overflowed cv2.remap's row limit, so the epipolar patches are now sampled in chunks. A segment that fit before is still one chunk. The test for this fails on alpha.

A3's failed 60% numeric criterion is now labeled a travel-axis proxy in README and METHODS. Wall-plane benefit is untested without wall annotations; all numerical tables remain unchanged.

#86: experiments: live dots, a macOS prototype of the holographic capture view

• Fog mask kept the previous view (r4115037778). An explicit seek counter resets the mask on same-keyframe seeks, forward/backward scrubs and resuming a settled preview. Natural same/adjacent playback and ordinary pause/resume preserve easing. The renderer uses the tested continuity decision for both motion settings; mode changes and fog re-entry still reset.
• Low-confidence depth carved lasting edges (r4115037781). Carving now skips pixels below Tuning.minimumConfidence, as fusion already did. The new test fails with alpha's DotField.swift.
• --still 2 --at 0.3 rendered keyframe 3 under keyframe 2's name (r4115037787). An offset past the keyframe's time on screen is now an error that suggests --still 3, because stills are named and compared by keyframe.
• Constellation hid every feature point (r4115916347). The edge birth time now applies only to non-feature sprites.

On b6bc4ad, swift build --jobs 2 succeeds and swift test --jobs 2 passes all 45 tests in seven suites. Both commands ran through run-guarded.py --min-free-gib 8. The first test build exposed direct mutating calls inside #expect; the tests now evaluate those calls before asserting their results. Production playback and reset decisions are covered, including same-keyframe seek, settled scrub resume under both motion settings and ordinary pause/resume continuity. Metal pixels and native interaction were not inspected.

#16: experiments: meter close-up reading, tap-to-confirm and two retake checks

• OCR joined distant fields on one row (r4114120264).
  • Boxes in a row now join only across a gap of at most 3 character widths. A character width is a box's width over its characters, in the image's own x units, so no aspect ratio is needed.
  • On the cached Vision reads (clean and locate runs, every configuration, strict and lenient), any limit from 2 to 6 widths leaves every number match unchanged, and 1.5 loses three. The committed results therefore stand.
  • I didn't rescore the sweep's degraded reads.
• A stale Keep survived a label change (r4114120267). Each saved answer now records the number it was given for, and the page keeps it only while it still shows that number. Answers saved before this change are dropped. The new test fails on alpha.

PYTHONPATH=src python -m pytest -q in experiments/meter-closeup/: 99 passed on the combined branch.

The deep-review follow-up found that CSV import discarded which reader label a Keep confirmed. Downloads and ingest now preserve shown plus whether it was recorded, including a reviewed empty label. Both the page and manifest reject unbound or mismatched imported Keeps. A legacy Keep cannot upgrade an earlier correction's verdict. Tests download, ingest, change the reader label and reload in fresh/stale browsers; numeric, empty and NONE labels, repeated ingest and prior corrections are covered. The original import fallback fails the new legacy-empty regression. Deep-review finding.

#124: tools: viewer that films a capture reaching the server, as a labeled replay or read-only live view

• Counters missed a final batch that couldn't animate (r4114688573). Count timing moved into count-timers.js. Each render queues what it drew and when its marks land. Its totals reach the counters once its own marks and every earlier render's marks have landed. A batch with no marks still shows, and counters never run ahead of marks in flight. Tests on a virtual clock fail with alpha's rule.
• Large batches expanded every file (r4115036581). markGroups chooses the at most six marks from per-group totals, making the same picks as before. send() reads nothing when motion is off.
• Run switching (r4114355831, r4114542142, r4114585684). No change. Commit 9c0680b made the status endpoint the only thing that chooses the run, and removed the ranking code the last two threads describe. An event naming a run the status hasn't named makes live.js read status at once (model.js:197, live.js:79, test at model.test.js:311).

node --test "test/*.test.js" in tools/integration-viewer/: 104 passed, including the headless browser test.

#134: docs: rebuild the README around the demo, with quick start, data and limits

• Animations ignored reduced motion (r4115037528). Both animated pictures now list still WebPs first, for each colour scheme, under prefers-reduced-motion: reduce. Each still is its animation's last frame, 29–37 KB. I checked the selection in headless Chrome against the live GitHub page and the new markup, under all four scheme and motion combinations.

Before (alpha, reduced motion)	After (reduced motion)

#14: docs: keep only the current plan and the facts that still hold

• Light-theme text contrast (r4112033407). docs/how-it-works.html gains text-only colours --phone-ink and --server-ink, while diagram fills keep the originals. In light mode, links go from 4.04:1 to 4.63:1, blue labels from 4.42 to 5.06, and orange labels from 3.20 to 5.03. Dark mode is unchanged and already at least 4.69:1.
• Coverage zone stopped at cable reach (r4112033409). The diagram now draws the candidate spots within the 20 ft of cable, and around them the zone the checks look at: about 11 ft out, past the cable's reach at both ends. Nothing outside cable reach is called "not needed". The caption gives each check's reach and calls the pool and driveway values placeholders.

Before (alpha)	After

Light theme at 900 px wide, from the page served locally. Dark-theme and 390 px captures were checked but aren't attached.

#18: Research handoff: live survey findings, options and tested prototypes

• The ground-replay patch loses evidence past a marked end (r4111600380). The defect is real. The patch is an unrun review artifact for an old baseline, so native-proposals/README.md now records it next to the patch's other known gaps, and the patch bytes are unchanged. Alpha's replay also drops regular sightings past current ends, but clearing a limit end calls replayPastLimits() for past-limit ground. It does not rebuild those regular wall/ground sightings. The README distinguishes these behaviors. The remaining replay loss is a fixable defect reassigned to the iOS caretaker, not a policy decision.

Verification accepts root scene.json or a scene under one shared top-level capture folder, including the public API Finder fixture. Unlike the public API's root precedence, it refuses both forms together so the verifier has one unambiguous input. Every non-metadata entry must lie beneath the shared folder for the wrapped form. All entries still count toward archive limits. Duplicate and ambiguous layouts remain rejected. Wrapped bundle review.

Merge order

Merge #240 first, then #242. #242 contains #240 at a7b3169 through merge commit daa36a2. Both include alpha 3c33ad7, retain the same runtime work meter and ingestion bounds, and preserve the shared Makefile dispatch repairs. Neither branch was rebased or force-pushed.

Still open

The ground-reach, variable-footprint and yaw contracts remain open. Photos-only uncertainty is an existing unsafe bound needing a product tradeoff; the iOS replay defect has been reassigned.

• How far out the worker reports ground. It stops at 10 ft, and the checks need the rule distance plus position error, 13 ft per the docs (#20 r4114022057).
• Variable battery depth in overhead sampling. The worker now samples both edges of the 0.56 m footprint, offset from the reconstructed wall front. The rear gap is fixed; other battery depths still require a server footprint contract (#20 r4114128923, r4115044069).
• The existing unsafe error bound on photos-only clearances. The current ±0.56 ft comes from surface pair error, not measured clearance-boundary uncertainty. Sam must choose a measured real error bar or withholding unsupported clearance evidence (#20 r4114128935, r4115044080).
• A per-capture yaw uncertainty for the meter anchor in packet 0.5 (#22 r4114003866).
• The fixable iOS replay defect in clearEnd, reassigned to the iOS caretaker (#18 r4111600380).

Checklist

• The scope is focused; related changes belong together. Every change answers an existing review thread.
• I explained what changed and why
• I included before/after images for visible changes
• The earlier static changes have before/after images above. This round adds tested fog-continuity repairs; no new native renderer video or runtime accessibility inspection was captured.

Current validation limits

Scene/rule derivatives and wall lookups are cached per judgment. Direct coverage builds no ground fallback strips, and redundant collinear vertices no longer multiply work. Full judgments at 13 and 20 subdivisions fail at the old limit and pass now within 20% of the one-segment representation. Distinct evidence IDs and duplicate diagnostics use hash lookups.

Metadata admission precedes schema validation, copies and resend comparison. It bounds walls to 50, vertices per wall to 500, runs to 30,000, checks and requests to 16,000 each, numeric leaves to 1,000,000, JSON nodes to 2,000,000, nesting to 32 and string/key characters to 4 MiB per document. Local case rules have 64 queries. Boundary controls verify clean admission and specific over-limit diagnostics.

Reported widths and corner allowances now share the same handling in both PRs. A case mismatch uses the response's reported value. Geometry adopts a validated reported override only for a different policy fingerprint. Matching public fingerprints retain the loaded rules and contract-check all reported battery dimensions, route height and corner allowance, including transformed responses. An unreported value is unverified when policy sources include private rules or the normalized rule fingerprint differs from the tested public rules, even if the assumed numeric value equals the public default. Four private-source/fingerprint controls fail on 30489f8 before this repair. Encrypted NPZ flags now trigger locked cache repair; the damaged-file control fails before the repair with a password-required RuntimeError.

Research-handoff has ten pre-existing Ruff lint diagnostics, confirmed against 6c0639e. Its formatting checks pass, and the formatting changes preserve Python syntax trees.

Python checks use the existing Python 3.12 environment at /Users/samgu/Programming Projects/house-scanning-real-intake-20261004/.venv/bin/python. Recon has 165 passing tests; its OpenCV 5.0 install is outside the declared <5 range. No dependencies were installed. Root Makefile dispatch cases pass. No model inference, dataset regeneration, private backend, simulator or phone checks ran in this caretaker round. Earlier experiment/viewer/media evidence above belongs to the original Opus run.

Summary by CodeRabbit

• New Features
  • Reduced-motion illustrations now use separate still images for light and dark themes.
  • Integration viewer counters track marks as animations land, including when updates arrive out of order.
• Bug Fixes
  • Live-dot visibility and fog transitions respond more consistently to playback changes and depth confidence.
  • Meter-number recognition is less likely to combine text that is too far apart, and saved review decisions are less likely to apply to a different number.
  • Free-space coverage now requires evidence from two camera positions with sufficient separation.
• Documentation
  • Clarified scan coverage boundaries and updated guidance on review, playback, and measurement results.

Null rules_assumed and null expectation query arrays now produce bad input records before counting. A regression saves both failed cases and a subsequent valid case in one report. The shared repair is merged from #240, which lands first. The matching-public wrong-width and malformed-case controls fail on a7b3169. Public-policy review. Case metadata review.

Repeated sweep and check expectation failures now produce one message per expectation, with a failure count, the reason once and at most three examples. A shared report budget retains at most 1,000 diagnostic entries and 262,144 serialized diagnostic characters, including duplicate JSON fields, and emits one omitted-count summary. Contract counts and acceptance are computed before display truncation; raw response payloads remain separate. Markdown-visible explanations receive budget priority over their duplicate JSON classifications. Both 1,000-error regressions retain the first explanation in the Markdown row and stay within the shared limits; both fail on c32ee97. Numeric cache and interval-map keys use exact rational strings, preserving equal integer/float values while preventing cheap modulus collisions. Regressions cover 1,000 runs with a 4,096-character reason, oversized reasons, JSON escaping and 2,000 colliding integers across four key paths.

The sampler count API now explicitly counts starts before the final endpoint; run_starts yields those starts and then the endpoint. Pinned db1e4c2 controls confirm that the old implementation already checked the last pre-end start, checked a sub-step run's first start and charged each yielded sample. The new count-API assertions fail there; the functional sampling controls already pass. Sampler review 1. Sampler review 2.

Free-space visibility now rejects camera rays blocked by fused occupied voxels, using the same stop-short tolerance as regular visibility. A nearer thin fused wall defeats both depth-backed rays, while clear rays and the target surface itself remain supported. The blocked-ray control fails on c557e3a. Fused-ray review.

[SamGu-NRX]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 30489f8a6c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Mark unreported private overrides as unverified

When a case assumes the public value and a server uses a private override but reports no corresponding value—for example, a wider private battery leaves no spot, nearest_considered, or width check—this condition suppresses the warning because configured[key] == value. judge() then runs expectations and geometry using the public rule and may claim contract_ok even though the response's policy.sources or rules_sha256 identifies different rules. Treat an unreported assumption as unverified whenever the response used non-public rules rather than using equality with the public default as evidence.

AGENTS.md reference: AGENTS.md:L33-L33

Useful? React with 👍 / 👎.

[SamGu-NRX]

daa36a2

[SamGu-NRX]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Nice work!

Reviewed commit: daa36a243b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[SamGu-NRX]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c557e3af38

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Sample every pre-end start in each sweep run

For any nonzero run, run_start_count already counts the sampled starts before the final endpoint, but run_starts iterates range(count - 1), dropping the last step (and for a run shorter than one step, dropping the first endpoint entirely). Coverage, footprint, and declared-height checks therefore skip valid battery positions and can accept a passing result whose omitted start would fail. Iterate all counted pre-end samples before appending b.

Useful? React with 👍 / 👎.

[SamGu-NRX]

run_start_count returns count + 1, including the final endpoint. run_starts therefore emits count - 1 pre-end samples and b. At c557e3a, [0, 0.1] with step 1 emits [0, 0.1], and [0, 2] with step 1 emits [0, 1, 2]; changing the range would duplicate or overshoot the endpoint.

[SamGu-NRX]

1ed45b6

[chatgpt-codex-connector]

Sample the first start before stepping through sweep runs

For any nonzero run, run_start_count already counts the sampled starts before the final endpoint, but run_starts iterates range(count - 1), dropping the last step (and for a run shorter than one step, dropping the first endpoint entirely). Coverage, footprint, and declared-height checks therefore skip valid battery positions and can accept a passing result whose omitted start would fail. Iterate all counted pre-end samples before appending b.

Useful? React with 👍 / 👎.

[SamGu-NRX]

run_start_count returns count + 1, including the final endpoint. run_starts therefore emits count - 1 pre-end samples and b. At c557e3a, [0, 0.1] with step 1 emits [0, 0.1], and [0, 2] with step 1 emits [0, 1, 2]; changing the range would duplicate or overshoot the endpoint.

[SamGu-NRX]

1ed45b6

[chatgpt-codex-connector]

Reject free-space rays blocked by fused occupied voxels

seen_through accepts a sample whenever the camera depth is at or behind the sample, but unlike seen_by it never calls vol.blocked (or otherwise tests occupied voxels between the camera and the sample). If the depth map sees through a thin or missed obstacle that the fused volume contains, two separated cameras can therefore mark the slab as observed and free_space can report clearance past the obstacle. Apply the same occupied-ray blocking test used by the regular visibility path before setting out.

Useful? React with 👍 / 👎.

[SamGu-NRX]

1ed45b6

[SamGu-NRX]

@coderabbitai review

[coderabbitai]

⚠️ Action not completed

Incremental review skipped.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[SamGu-NRX]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 👍

Reviewed commit: 1ed45b65b5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[SamGu-NRX]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: c4461b22bb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".