Harden v0.2.3 clarity and artifact safety - #15

Changes

• redraw the generated benchmark SVG around measured 0% negative controls and continuation-only timing; scope the title and strict-success label to the saturated v0.2 calculator task, shorten the diagnostic badges, and remove the unsourced ~555 s figure
• generate protocol-v2 examples from each environment's active ToolSpec, with built-in/custom round trips and byte-frozen protocol-v1 fixtures
• persist entries_per_shard in the cache index and allocate new shards after the highest numeric suffix found in either the index or cache directory
• make pruning transactional: atomically publish a copied next index before best-effort deletion, preserving every old shard if publication fails and tolerating harmless orphan shards if cleanup fails
• reject unverifiable schema-v1 caches, atomically publish cache shards/indexes, canonicalize tokenizer identity inputs, and clean up failed model construction reservations
• tighten the dark banner and bilingual onboarding, and correct tokenizer and CUDA-install documentation without expanding either README

Why

The published figure needed task scope and had an unsourced teacher-preparation duration. Cache reopen could also reuse a surviving shard suffix after pruning, lose its configured shard size, and delete old shards before the replacement index was safely published.

The new regression writes shards 0-3, prunes 0-1, reopens, writes shard 4, and proves shards 2/3 remain byte-identical and readable. Separate fault injections prove failed index publication leaves the old index and all old shards intact, while post-publication deletion failure only leaves an unreferenced file.

Impact

No algorithm, environment, benchmark, model family, or GPU experiment was added. The v0.2.0/v0.2.1/v0.2.2 tags, public adapter revision, and frozen benchmark JSON remain unchanged. Version remains 0.2.3.dev0; this PR does not publish v0.2.3.

Validation

• ruff check . passed
• ruff format --check . passed
• mypy src/miniverl passed
• pytest -q -m "not gpu and not network" --cov=miniverl --cov-report=term-missing --cov-fail-under=80: 1085 passed, 87.30% coverage
• pytest -q -m gpu: 5 passed
• pytest -q -m network: 3 passed
• package build and twine check: wheel and sdist passed
• 92 relative Markdown targets across 40 tracked files passed; changed English docs/changelog passed external link checking
• generated SVG byte comparison passed; rendered and inspected at 1120 px and 820 px
• frozen JSON SHA-256: 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc