Re-land #8055 with data-initializer roots: classify test data declarations instead of refusing selection

[gunbai-bot]

DRAFT — INCOMPLETE. The branch in its current state still reds main. Do not merge, do not flip to ready.

Re-land of #8055, which #8104 reverted at ca64dbae.

Why #8055 was reverted

It made live-read classification load-bearing at every selection decision, but the classifier's subject universe is fn-arrow declarations while the enrolled population contains 38 test data rows across 14 files (census re-measured on this branch, independently matching the reported figure). Those rows cannot bind, so the manifest carries a Refused row and the consumer stops the line:

AFFECTED-SET REFUSAL cause=LiveReadClassificationRefused

The refusal itself is correct and is not weakened here — it exists so a modelled refusal cannot be folded into a boolean and pass as SelectionSuperset. The defect is that #8055 made classification load-bearing before teaching it the test data form.

State of this branch

#8055 restored by cherry-pick of 7f7da933	done
#8099's P3 cohort probe	deliberately not restored — measurement plumbing for the implementation, not needed to establish behaviour
main's ci_layer_roots #7446 citation correction	verified present after the pick
gate_test note	took #8055's wording; both sides carry the fixed symbol, they differ only by a G1 qualifier that is wrong once G2 returns
the test data repair	NOT DONE — this is the whole point of the branch
ci.yml	to be regenerated from authorities before ready, never hand-resolved (#8104's own lesson)

The repair, as designed so far

Feasibility established rather than assumed: the lens consumes a host-marshalled FnArrowDecl stream (v2.std.fn_index fn_arrow_decl_facts_live, itself a registered Scaffold), and data declarations are reachable in the Node tree — v2.lens.mandatory_tag already reads ^dag_surface_data_decl_name / _type / _expr, and _expr is the initializer body a data root classifies from. So this is a marshal extension plus a second root shape, not a new parser and not a cast of a data initializer into FnArrowDecl.

Binding must be decided once over the combined population. bind_g2_root's existing law exists because zero and many both looked like success: zero matching declarations seeded an empty walk that reaches no carrier and therefore classified LocalRead — the strongest answer, the one that licenses a skip. Trying fn-roots then data-roots and taking the first match would reintroduce that exact empty-observation narrow one level up. So: zero → Refused(Unbound), exactly one → classify that form, many → Refused(Ambiguous), over both populations at once.

Open, and not mine to decide

Whether a known-but-unsupported root form should force that entry to run (typed, counted, per-row) rather than refusing the entire affected-set computation. That is a three-way decision — classified / conservative-run / refused — reserving global refusal for genuine substrate incoherence. It is likely right and it changes what a refusal means on a load-bearing path, so it is awaiting operator sign-off rather than being inferred.

Acceptance bar (behaviour first, no falsifier run required)

• all 38 test data roots bind and classify; unsupported = 0, ambiguous = 0, missing = 0
• a paired control: a test data root classifies, AND a genuinely unclassifiable root still refuses and is counted — the first alone is satisfiable by weakening the refusal
• the incident controls unchanged: total_entry_groups = 2, selected = 1, skipped = 1, B skipping via skip-before-resolve-fast-path
• both G1 restoration mutations still produce 2 selected / 1 expected

No budget change of any kind is part of this branch.