feat(tutor-quality): add stage 2b semantic foundation

[ttbombadil]

Summary

Implements Tutor Quality Stage 2B PR 1 / Milestone 1: the deterministic
semantic-evaluation foundation below the future Judge layer.

It introduces:

• the Stage-2B-owned Semantic Corpus
• separately versioned candidate-specific CaseExposureRecords
• Frozen Pre-Turn Context validation
• the narrow Stage-2A Case Adapter
• Stage-B-owned transcript references and canonical transcript digests
• evidence-backed compatibility validation for existing Stage-2A transcripts

TQ-SEM-001 is the only executable Semantic Case in this PR.

Architecture

• Stage 1 remains unchanged.
• Stage 2A remains the sole Tutor execution path.
• runTutorQualityEvaluation remains unchanged.
• the Stage-2A Anchor Corpus remains unchanged.
• the Stage-2A Case Adapter is pure mapping only.
• candidate-specific exposure history is separate from Semantic Corpus
  versioning.
• Stage B computes its transcript digest without modifying Stage-A artifacts.
• compatible existing Stage-A transcripts can be ingested even when
  Stage-A scenario IDs differ from Semantic Case IDs, but only when full
  Frozen-Context compatibility is proven.
• no fuzzy transcript matching is used.

TQ-SEM-001

Adds the INPUT_PULLUP development case with:

• exact Frozen Pre-Turn question/answer binding
• a draft sketch-grounded factual reference, explicitly marked pending review
• no Gold claim
• no calibration claim
• no held-out claim

The factual reference does not generalize LED_BUILTIN to pin 13 without
declared board evidence.

TQ-SEM-002 through TQ-SEM-005 remain gated by DG-14 and are not executable.

Explicitly out of scope

• Semantic Judge
• StructuredLLMProvider
• Tutor/Judge provider calls
• calibration
• repeated model benchmarking
• pairwise evaluation
• ranking/statistics
• Stage 2C

Local verification

Node 24.20.0:

• npm run check — passed
• npm run test:tutor-quality — passed, 263 tests
• npm run test:unit — passed, 2524 passed / 1 skipped
• npm run check:docs — passed, 51 Markdown files
• ESLint — passed
• git diff --check — passed
• ./run-tests.sh — completed successfully
• Docker — 26/26
• Playwright E2E — 25/25

SonarQube:

• Quality Gate — PASSED
• New-code violations — 0
• New-code coverage — 82.4 %
• Duplicated lines density — 0.19415 %
• Security hotspots reviewed — 100 %

Knip

• npx knip reports 0 findings.
• The baseline-aware Knip gate reports:
  new: 0, baseline: 0, resolved: 0.
• PR 1 introduces no Knip diagnostics.

No Sonar rule was disabled, no suppression was added, and no verification hook
was bypassed.