[M2-ISO-02] Budget revision: re-baseline iso_depthkey_rebuild to 0.3 ms (zero-margin CI gate)

[offdev]

Why

CI is red on master (run 37194767648) and has flapped on PR lanes: the
iso_depthkey_rebuild budget gate (PRD §8.1, mean ≤ 0.2 ms for 10k
dirty cells after a terrain edit) fails the Linux x64 (clang++)
reference lane on unchanged engine code.

The gate's platform is the CI reference machine (methodology §5:
"CI is the gate"): ubuntu-24.04, Clang 18.1.3, CMake Debug, shared
runner. That toolchain measures this workload at 0.194–0.267 ms —
straddling the 0.2 ms bar. Whether the gate passes is decided by how
busy the shared runner is at the moment, not by the engine:

CI reference-lane run	fpx16_16 mean	fp32_pinned mean	result
37192995927 attempt 1 (PR #72 lane)	0.194142	0.201495	PASS / FAIL
37192995927 attempt 2 (slow runner)	0.266709	0.267328	FAIL / FAIL
37194767648 (master, commit 70830a7)	0.202146 / 0.202204	0.202018 / 0.194168	FAIL (one backend passed 0.07% under the bar in the same job)

Not a regression: the engine's setTile path is unchanged since the
2026-10-01 division-free workload fix (five merges in between, none
touching iso_depth_table), and the Linux x64 (g++) lane of run
37194767648 passed the same workload on the same commit.

The 0.2 ms bar was calibrated from local-machine runs (0.081 ms
canonical g++ -O0; 0.139 ms local Clang -O0) and was never re-verified
with margin on the reference toolchain — the exact lesson the
m2-iso-depth-table-workload-fix baseline recorded.

The fix (PRD revision, per the NFR-8.1 policy)

The PRD §8.1 policy says a budget "is revised via a PRD revision" — so
this PR does exactly that, in one change:

• PRD.md — v0.4: §8.1 row ≤ 0.2 ms → ≤ 0.3 ms (+ change note).
• budgets.json — target 0.2 → 0.3; measured 0.0814067 →
  0.202204 (latest recorded CI reference value, worse backend).
• New baseline docs/benchmarks/baselines/m2-iso-depth-table-budget-rebaseline.md
  (eighth baseline) — verbatim CI reports, before/after, margin math:
  0.3 ms is 48% above the latest recorded reference value and 12.4%
  above the worst observed reference-lane mean (0.267328, slow-runner
  outlier). The gate's metric is the mean (n=3000, warmup 100), so
  single-iteration spikes don't drive it.
• Docs — docs/api/iso_depth_table.md, docs/api/iso_depth_key.md,
  docs/concepts/coordinates.md, docs/README.md, the baselines
  index, and the M2/M5/M8 roadmap budget references updated in the
  same change (DOC-003); milestone change-log row added.

No engine code, workload, or test changes — this is a
calibration of the gate's margin on the reference toolchain, not a
benchmark change to hide a regression (methodology §1; no regression
occurred). 0.3 ms for 10k dirty cells is 1.8% of the 16.7 ms frame
budget — the product promise stays tight, and a future >10% regression
still trips the 10% band against before = 0.202204.

Verification

• LAIGE_BUDGETS_PATH=$PWD/budgets.json ./build/bin/laige-render_tests --gtest_filter='IsoDepthTableBudget*' — both backends PASS
  (after≈0.082/0.084 ms local, before=0.202204, target=0.3).
• laige-core_tests --gtest_filter='BudgetHarnessTable*' — 4/4 PASS
  (repo budgets.json still schema-valid, 16-entry pin intact).
• CI on this PR (Linux g++/clang++ reference lanes run the gate
  against the new 0.3 ms target) is the definitive check.

Note: the failed master run 37194767648 was re-run (attempt 2) while
this PR was in preparation; once this PR merges, the next master run
carries the 0.3 ms target.