The bare TerrainGenConfig defaults did NOT reproduce the terrain the developer
locked, so "run the default generator" was not "the terrain in the gallery" —
the single most expensive fact in the codebase, and the reason a fresh chat
would spend an afternoon chasing differences that were CONFIGURATION, not
regression. This is the deliberate task that ends that, before any river work.
A1 — the defaults ARE the locked shape now. Five fields actually move:
SpeckRevert false->true, MinLandComponentFrac 3e-5->2.5e-7 (120x smaller; the
config default would have eaten real islands, not specks), SouthStretch 0->2,
FragmentAmp 0->0.5, Erosion false->true. Seven more were already correct via
SouthernStretch.Default* / CoastalFragment.Default* and are now pinned as
literals, because TerrainShapeV1 used to do that pinning and this default set
inherits the job. CoastShelf stays OFF — the locked shape has no shelf, and
evaluating it (D-041) is its own later task once water renders. Offshore stays
Off permanently (D-063): islands are organic-only, made by the stretch +
fragmentation and identified by the region layer, never placed.
A2 — the preserve mechanism. The curve knots are percentiles of the FAMILY-OFF
land distribution; flipping the defaults would have moved the pool, the knots,
and with them the render field of every batch including terrain-shape-v1
itself. So the pool is pinned family-off (TerrainGenConfig.WithFamilyOff /
CalibrationPool) rather than the knots being baked: calibration stays live, its
INPUT distribution is held still. The pin was a no-op by construction — it sets
the values the defaults carried the instant before the flip — and re-measuring
after confirms it: pool, all six knots, per-seed spread, shaped max,
monotonicity spikeMax and all seven band shares identical.
Applied wider than "in CalibrateCurve": OffshoreIslandsTool,
RegionLabelingTool and SouthernStretchTool generate their own family-off
field for the Phase-1 anchor, so the pool pin alone would NOT have covered
them and their a1 would have failed for a configuration reason. TerrainGenTool
too — it AUTHORED 02_pass1_port and must stay able to regenerate its own
anchor.
Recorded as a judged-and-parked property: knots measured family-off, applied
family-on. Deliberate, not an oversight. Same disposition as the mid-slope
feather.
A4 — no oracle may pass against a superseded baseline. Six anchors retired
(01/03/04/06/08/09) with their checks and ISLA_T0x_SOURCE defaults; three kept
(chat1/02_pass1_port as the family-off pass-1 guard, chat2/10 and chat2/11 as
the shape and erosion acceptance anchors). Two invariants were RE-POINTED
rather than lost — the southern stretch's north-lock and the coastal-fragment
interior-lock now compare against SAME-RUN fields, which is scale-free and
cannot be invalidated by a moved dump. The retired dumps are kept, not deleted,
and marked superseded in their INDEX.md.
A missing anchor is now LOUD. The old pattern skipped silently, so a moved
anchor did not make its oracle fail — it made it not RUN, and a batch with a
skipped check prints an all-PASS table that reads like a clean one. That is
the INVERSE of the hazard the re-baseline guards against, and the migration
below is exactly the event that would have triggered it, on nine anchors at
once. ShapingOracle.LoadAnchor now separates the two cases: absent -> throw;
present at another size -> loud INCONCLUSIVE, which is a fail, never a pass.
TerrainShapeV1 inverted from PRESET to GUARD and moved to its own file.
Apply() is gone — stamping the values on top of the defaults would MASK a
drift instead of catching it. Its constants are now the assertion target, and
Assert() refuses a run whose defaults have drifted off the locked shape.
B/C — batches are namespaced by chat: batches/<chat>/NN_slug/. Task numbers
restart at 00 per chat, so a flat root collided the moment a second chat
existed — four colliding prefixes across 25 batches, separable only by slug.
ToolingPaths.ChatSlug is REQUIRED (throws if unset) and defaults per tool to
its authoring chat, so re-running reproduces a batch in place while ISLA_CHAT
redirects — which is also what stops an acceptance run from overwriting the
very anchor it checks against. Writes go through BatchRoot; historical READS
compose against BatchesRoot and so carry the prefix in their own source string
("chat1/02_pass1_port"). The 25 existing batches were migrated moves-only.
ACCEPTANCE — 16 of 16 byte-identical, 0 failed. All 8 gallery seeds at 8192
from the bare defaults are byte-identical to chat2/10_frag4_seed_gallery
(= terrain-shape-v1, a59e52f); all 8 erosion fields byte-identical to
chat2/11_erosion (= ea291ea). Every gallery table row and every erosion
statistic reproduces its recorded value exactly. DrainageTool's a11 passes
bit-identical over 67,108,864 cells, and its analysis reproduces batch 12
exactly — so the whole chain rivers depends on (shape -> erosion -> drainage)
is unchanged. All 12 edited tools re-run clean; both new guards negative-tested.
The baseline moved in DEFAULTS, not in TERRAIN.
-> XX_Human/output/rivers/01_rebaseline_and_batch_namespace.report.md
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WhkXBQh2tDmcWKpXYcj8vj
chat2/02 dissolved the terraces and lost two thirds of the mountain with them, then
concluded the loss was structural and needed a Phase-1 noise change. That conclusion
was wrong, and this commit is the refutation.
A monotone curve is a free reparametrization: it may be gentle in one place and steep
in another, and can lift bottom-heavy input into a substantial massif without ever
going flat. The area above a height is set by where the percentile->height mapping
crosses it, and that mapping is entirely ours to choose. The 02 sweep that "proved"
the loss structural varied climbFeather, which shapes the JOIN, not the mass
distribution — the wrong knob, and too strong a conclusion drawn from it.
Core/ClimbCalibration — the climb's control points are now MEASURED off the staircase
instead of invented from shape knobs. For p in {10,30,50,70,85,95} of above-ceiling
land, take that percentile's raw height and its staircase output height; PCHIP through
the pairs. That reproduces the staircase's elevation envelope, so the mountain mass
returns, while MinNormalizedSecant floors every grade so the flat bench and plateau
interiors become slope. The floor bites on exactly one segment — the plateau — which
is precisely where the staircase was flat.
ContinuousCurve.BuildCalibrated joins it to the same pinned lowland handover, the same
C1 join and the same per-seed spikeMax. The 02 analytic path survives unchanged as the
"before" contrast, and deliberately keeps its strictly-increasing-secant rule: a
calibrated curve is WAVY by design, so convexity is the wrong invariant for it and the
secant floor is the right one.
peakSharpness replaces summitDrama and fixes its bad trade. Drama steepened the peak by
pulling the summit ONSET down, dragging the whole massif with it (p99 199 -> 121 m).
Sharpness reshapes only above the last measured percentile, leaving that height fixed,
so peak and massif are independent: raising it leaves p90, >100 m and >220 m untouched
and only moves land within the summit.
Measured, both seeds, 2048:
variant >100 m >220 m p90
staircase (target) 16.05% 4.53% 127.4 m
continuous_02default 4.80% 0.63% 58.4 m
continuous_restored 14.25% 3.45% 123.2 m
continuous_bigger 19.70% 5.80% 163.5 m
Oracle all hard checks pass, including (a2) staircase still bit-identical to task 01's
dump and (d) lowlands bit-identical across every calibrated variant. New (g) reports
land above 100/220 m per variant and is deliberately NOT gated — it is a taste target
the developer tunes, and gating it would make mountainLift unusable. What it must never
do is stay silent, which is how 02 lost the mountain unnoticed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DCWNaDZPfTiAy3meGNGgqt
