- new named constant LANDMARK_I32_WEIGHT_CEILING (0x7fffffff) in
landmark.ts: a seconds/quantumS quotient at or past 2^31 would wrap
NEGATIVE on the Int32Array store, silently bypassing the
LANDMARK_MAX_QUANTA throw in dijkstraPass (a negative contribution
never exceeds it) and — at a -1 wrap — becoming the 0xFFFF sentinel;
prepare() now fails loud at the table-build entry, naming the edge,
the ceiling and the fix (raise landmarkQuantumS)
- 'nothing is clamped, ever' now holds at the storage boundary;
validation is per-edge at build entry — the search hot path stays
allocation-free and validation-free
- L5b regression vector: 3e10 s edge (3e11 quanta) throws in both
landmarkTables and farthestPointAnchors; just-under-ceiling
(2147483646) still fails via the documented 65000 law; ordinary
weights untouched
Debug session alt-potential-4x-slowdown, root cause RC-1 (LIVE correctness
bug): bound()'s per-node sentinel skip (tv === 0xffff -> drop landmark l)
varied the active landmark family set PER NODE, so the toTable max lost
cross-edge consistency across edges leaving the 6,285-node disconnected
island (2.02% of cells) — shipped routeAstar returned suboptimal seconds
on 1/194 fresh random real-graph pairs vs the Dijkstra oracle (worst
single-edge family variance 1,026 s). D-03's fixed corpus was green on
sampling luck.
The fix (reverseFor precedent, in-constraint):
- per-graph WeakMap memo cleanColumnsFor: clean[l] = 'to-column l has ZERO
sentinel cells' — O(L*n) sequential row scans once per graph/isolate
- per-search active list activeL = clean columns only; bound() loops the
active list with NO per-node sentinel branch; the constant-table
goal/origin NaN checks are retained (belt)
- on the current binary (0/24 clean columns) potentials degrade to EXACTLY
haversine — the designed sound fallback; the pre-ALT CPU band returns
- ALT stays live and tested on strongly connected fixtures (mini-graph
0/2 dirty columns, A12 ring all-clean, new A13 mixed subset case)
- LANDMARK_SENTINEL_U16 named constant; sentinel compares at setup scope
only; pure-module law untouched
BH stop rule, reductions, balanced alternation, meet probes, real-seconds
reconstruction: UNTOUCHED. F2 (transposed table) explicitly out of scope.
Battery: routing 81/81 (A11 RED pins now green, A13 new), bot 126/126,
routing+bot typecheck clean.
Task 1: the D-03 gate script (measure-optimality-gate.mjs) reads the frozen
pre-ALT record, re-routes the same coordinate-pinned pairs on the current
ARG2 binary, and gates on seconds-equality (±0.1%) + verdict-class
preservation — never on node paths (equal-cost ties are documented behavior).
Rule-1 deviation the gate forced on its first run: routeAstar returned
SUBOPTIMAL real-graph routes (x038: 1234.6s vs the 1214.8s Dijkstra optimum;
x045: 1475.5s vs 1460.6s) — the fixture sweep could not see it. Root cause:
the fromTable bound d̂(ℓ,t) − d̂(ℓ,v) is one-directional Lipschitz (its u-side
term needs the REVERSE edge; a oneway does not provide one — the anchor
reached node 264611 126.8s faster than its own 1.49s upstream neighbour
264610), and the shared Ikeda average p = (h_f − h_b)/2 is consistent only
when every family inside h_f/h_b is two-sided. The toTable bound
d̂(v,ℓ) − d̂(t,ℓ) rides stored traversable edges in BOTH directions; it alone
joins the haversine now. Stop rule, reductions, meet probes, reconstruction:
untouched. landmarkTables/build verified EXACT (0 mismatches vs independent
integer Dijkstras, both rows, all 24 anchors).
- measure-optimality-gate.mjs: exported core + CLI guard (route-corpus
pattern); epsilon mirror + argVersion asserts precede any measurement;
per-pair rows, extract-drift WARNING (never a re-record), honest exit law
- astar.ts: potential() consults landmarkTo only; doc-comments carry the
one-directional-Lipschitz law and the x038/x045 evidence
- astar.test.mjs: A11 re-pinned (landmarkFrom must NOT appear in the
comment-stripped search — re-adding it trips the pin); A12 oneway-ring
sweep added (directed reach asymmetry stays Dijkstra-equal on every pair)
Battery: routing 80/80, bot 126/126, both typechecks green.
D-03 gate: GATE PASS — 53/53 routed pairs equal within ±0.1% (max |Δ| 0.005%),
verdict classes preserved (6 unreachable + 41 unsnappable, 0 re-snapped).
- routeAstar's ONLY change: the h_f/h_b bodies become max(ALT triangle
bounds, haversine) — per-search Float64Array(L) constant tables in quanta
(0xFFFF written as NaN = search-scoped landmark drop), lazy per-node
max over the four triangle families with RAW u16 sentinel compares
BEFORE any float conversion, haversine max-fallback on sentinel-flooded
nodes (Pattern 3 containment)
- the (h_f - h_b)/2 combination, reductions, balanced alternation, meet
probes, BH stop rule, and real-seconds reconstruction are byte-identical
(research Pattern 1 — the survey's (pf-pr)/2 is what p_f already is)
- no Map/Object allocations (A8 green), no new imports, one search path
(L >= 1 is the decodeGraph guarantee, D-01)
- A1's 625-pair Dijkstra-equivalence sweep re-pinned under ALT: the
standing optimality proof; A9 island + A10 skip laws sealed; A11
structural pin green
- module header + routeAstar doc gain the ALT paragraph (c-metric
exactness, sentinel laws, Ikeda history intact)
graph.ts versioned IN PLACE (no second codec): GRAPH_MAGIC 0x41524732,
GRAPH_VERSION 2, HEADER_SIZE 72 (72%8==0 keeps f64 views aligned);
header gains landmarkCount u32 @60 (the old pad), landmarkQuantumS f32
@64, reserved u32 @68; the landmark suffix (u16 toTable, u16 fromTable,
u32 landmarkNodes) extends layoutFor's total arithmetic — exact-size
law covers the variable suffix BEFORE any view allocation (T-08-01).
D-01 fail-honest: ARG1 bytes / version!=2 / landmarkCount=0 / hostile
quantum / size mismatch all decode null (committed mini-graph-arg1.bin
is the standing rejection vector). GraphInput.landmarks REQUIRED and
loud-validated (T-08-04). RoadGraph gains the five landmark fields
(Plan 08-03's ALT potentials consume them).
Fixture regenerated as ARG2 through the SAME encode authority: the
builder gains the tiny real landmark stage (L=2, anchors [0,18],
quantum 0.1s, coverage 1.0 — identical bytes every re-run).
Rule-3 call-site updates forced by the required-landmarks law (missed
by the plan's files list): astar.test A3 + snap.test S3 synthetics
carry inert hand landmark blocks. Fixture regen pulled forward from
task 3 so this commit's own battery is green (74 passed, 0 failed).
farthestPointAnchors + landmarkTables + landmarkCoverage over the
ĉ metric: edge weights floor-quantized once (Math.floor(seconds/q)),
exact integer Dijkstras (Int32 labels, binary heap, lazy deletion),
0xFFFF sentinel, 65000-quantum loud overflow. Traversal law mirrors
astar.ts verbatim: bit0 edges never relaxed in EITHER direction.
Test-vector corrections caught at GREEN (expectations were wrong,
implementation was right):
- L4: rows index by anchor POSITION (row-major [l*N+v]), not node id
- L6: the c-metric farthest node from corner 0 is interior node 18
(297q: interior hops cost 2x perimeter), not corner 24 (264q) —
verified against an independent brute-force Dijkstra
- upload-road-graph.mjs: wrangler-shelled maintainer tool (pre-flight 1-80MB,
round-trip verify hint); bucket araharah-graphs NOT created — R2 is not yet
enabled on the account (wrangler code 10042) — dashboard one-time activation
is the phase's user_setup item
- measure-route-cpu.mjs: 25-pair gate (corpus + seeded quadrant spread),
hrtime around routeAstar, warm-up excluded, unreachable probes reported
SEPARATELY from the routed p95, exits 1 over the 8ms budget
- Rule 1 (measured): the class set lacked *_link ramps — the graph FRAGMENTED
(7.7% of nodes in severed motorway islands; corpus pairs 'unreachable').
5 link classes added -> 311325 nodes / 677068 edges / 14.7 MB, all landmarks
in one component; enforced floor corrected 1KB -> 1MB (D-01)
- ladder (a) drop tertiary+residential: p95 17.8ms but 12/25 pairs unreachable
— REJECTED on correctness (D-03), reverted, ladder history recorded
- ladder (b) bidirectional A* (BH, Ikeda potentials, reverse-CSR WeakMap):
provably optimal (A1 625-pair sweep green), alternation balanced by heap
size (raw-top comparison starves the backward side 19501:3), reduction
signs in the standard feasible form. Measured: p95 ~38-93ms — NO GATE PASS.
The potential split halves each side's steering, cancelling the halved radius
- GATE FAIL is the honest verdict: ladder (c) — Workers Paid vs layout-v2 ALT
(still $0) vs free-plan-with-degradation — requires maintainer sign-off;
docs/ROAD-GRAPH.md carries the decision table and the fork contract
- bot/src/routing/worker-router.ts: module-scope per-env graph memo (1 R2 GET per
cold isolate, shared with the 06-07 cron arm via loadGraph), snap+A*+deriveSteps,
claim-gated 401 / 4096-cap 413 / V5 bbox 400, 200-null on every miss class,
ONE best route (no alternatives), category-only logs
- webhook.ts: /api/route arm delegating to handleRoute (405 non-POST)
- config.ts: ROAD_GRAPH: R2Bucket — first R2 binding (never a KV namespace, GE5)
- wrangler.jsonc: r2_buckets ROAD_GRAPH -> araharah-graphs with the fork law comment
- routing/src/http-source.ts: createHttpRoutingSource — slug-guard-first, fail-honest
(throw/non-200/malformed/null-body -> null), defensive RouteResult validation;
exported from index.ts; ladder note: backend #2 LIVE
- A7 re-pinned (06-01 named 06-04 as the owner of createHttpRoutingSource)
- test-side fix during GREEN: the R2 fake hands decodeGraph an exact ArrayBuffer
(Node Buffers are views; encodeGraph output is already an ArrayBuffer)
- bot 103 passed / routing 65 passed; both typechecks clean
- snapToGraph: bbox+margin guard first (honest null = router-pending,
D-03), then 3x3 coarse-grid neighborhood (0.002 deg cells, WeakMap
memo per RoadGraph — LAW 1 lazy memoize, O(local) not O(n))
- projection in cos(lat)-scaled degree space (Pitfall 8); returned
distance is haversineM; result carries the best segment's nameId
(null for the unnamed sentinel) for plan 07's cron street-hop
- oneway-reverse edges stay snap-visible (direction is A*'s law);
gridBuildCount() instrumentation kept internal to the workspace
- standard encoded-polyline algorithm at 1e6 precision, double math
(no int32 shifts) so no coordinate range wraps silently
- decode THROWS on truncated varints — never partial/NaN points (T-05-02)
- round-trip law encode(decode(s)) === s proven on all vectors incl.
the external live-OSRM geometry (RESEARCH Finding 1)
araharah-executor