[mache-eb9b30] design(adr): ADR-0019 real-corpus fixture registry (2-tier, fixed-anchor baselines)

[jamestexas]

Summary

ADR proposing a real-only, snapshot-in-tree, 2-tier fixture registry as first-class test infrastructure. Frames the "real corpus catches what synth misses" pattern from PR #404 (the dead_code 460× speedup) as structural substrate rather than ad-hoc heroics.

Decision (post-assessor amendments)

• Snapshot in-tree at pinned SHAs (rejected: clone-on-demand, env-var live-pointer)
• 2 tiers, ALL real: medium (100-500 files, matrix runner default) + large (1000+, perf gates). No synth. No tiny.
• Fixed-anchor baselines + tolerance band (rejected: auto-rebaseline — it bakes in a perf ratchet)
• ART-owned snapshots only for v1 — external/MIT fixtures deferred to license-policy ADR
• 3-PR migration (registry API → external snapshots → baselines), each independently shippable

Why the assessor pivots

Original draft red-teamed by paradigm-assessor; 1 fatal + 4 significant findings:

Finding	Resolution
Fatal: Auto-rebaseline mathematically bakes in invisible perf rot. At 25% tolerance, 4 merges of +20% = 2.07× slower with no gate firing (1.20⁴).	Replaced with fixed-anchor + explicit task fixtures:rebaseline with audit trail
Tiny tier can't catch the bug class the ADR opens with. Real micro-libs (xxhash, etc.) have ~5 public functions; PR #404's bug was OR-join blowup on thousands of nodes — structurally unreproducible at tiny scale.	Dropped. Registry starts at medium. Sub-medium tier deferred to future ADR if real need emerges.
Size estimate undersold 5-10× (rosary full tree is ~125 MB, not the claimed ~5 MB).	Added curation spec in D.7. Real v1 measurement: ~15 MB.
License framing one-liner for a categorically new compliance surface (attribution, supply-chain, distribution).	Promoted to its own Consequences subsection; v1 restricted to ART-owned snapshots.
6-PR bundling-as-ambition (same shape as ADR-0018's original draft).	Reduced to 3 PRs, each independently scoped.

Full assessor log: _agent_log/paradigm-assessor_2026-05-19_audit_log_adr0019.md

What this ADR does NOT decide (deferred to follow-up ADRs)

• External non-ART fixtures + license policy
• Large-tier sourcing (kubernetes/pkg subset, etc.)
• Sub-medium fixture story (if a real need emerges)
• Rolling-median observability dashboard (non-gating, layered on baselines)

Each becomes its own ADR when there's evidence to debate it against.

Test plan

• ADR red-teamed by paradigm-assessor; substantive concerns addressed
• CI passes (docs-only PR; no code)
• Reviewer agrees with scope + assessor pivots

Beads

Tracking: `mache-eb9b30`
Supersedes: ADR-0017 SB-05 (synthetic-medium generator — explicitly rejected per user direction)
Pairs with: `mache-655e98` (rough-edges parity gate), ADR-0017, ADR-0018