Wire the absolute per-cell trainability gate into scripts/issue2225_train.py (the sibling entrypoint #2242 left ungated)

Goal

Close the mechanical arm of the #2242 absolute per-cell trainability floor on the ONE remaining main-resident fine-tune entrypoint that can still train a structurally-untrainable cell without any gate firing: scripts/issue2225_train.py.

Why this exists

#2242 installed the absolute floor (below it a cell is DROPPED, not shrunk to 1 row) across four surfaces:

• D1 — the shared gate in src/explore_persona_space/artifacts/datagen.py (trainability_floor_rows, assert_cell_trainable, CellTrainabilityError, and an additive min_rows_absolute entry assert on generate_training_data).
• D11 — the one UNCONDITIONAL runtime arm, in scripts/issue778_finetune.py at the load_dataset row-count site, before the smoke slice and the model load.
• D2/D4/D5/D6 — the review-side arms (on-policy-completions.md, planner-section-reference.md, critic-lens-reference.md, planner.md), which bind every FUTURE plan through the planner + critic.
• D7/D8/D9/D9-bis/D10 — the lint surface pin + tests.

D11 was deliberately scoped to the incident's own script (#2221's demonstrated reuse magnet). The #2242 code review's bug-class sweep then found the sibling: scripts/issue2225_train.py is a main-resident fine-tune entrypoint that extends the same recipe — it has its own load_dataset + TRL SFTTrainer and imports the same module constants from issue778_finetune.py (ft.PER_DEVICE_BATCH, ft.LORA_R, …) — and carries NO trainability gate.

That review recorded it as a standing recommendation, explicitly NOT a condition on the #2242 round: #2225's own runs are complete, and any future plan reusing its trainer is now bound by the D4/D5 review arms. This task is that recommendation, filed rather than parked as a chat note (CLAUDE.md § Workflow-fix-on-bug protocol: a concrete workflow-surface suggestion in an agent's report prose gets the same auto-file + spawn as a formal workflow-fix-candidate block).

Scope

1. Add the gate call at scripts/issue2225_train.py's row-count site — the load_dataset result, BEFORE any smoke/subset slice and BEFORE the model load, so it sees the FULL realized row count and raises before any model download, GPU allocation, or trainer construction. Mirror the D11 shape in scripts/issue778_finetune.py (its _gate_cell_trainability helper is the template — reuse it directly if importable, rather than writing a third copy; check first, per the reuse-discovery rule).
2. Derive effective_batch_size and epochs from THIS script's own constants, not #778's — the two scripts share constant NAMES via import but a divergent value would silently mis-derive the floor. State the resolved arithmetic in the failure message the way D11 does.
3. Thread the same two override flags (--trainability-floor-override + --trainability-override-reason), preserving the invariant that an override without a reason raises. If #2225 has a wave/fan-out dispatch path, thread them through it too.
4. Smoke demotion: on_fail="warn" under this script's own smoke discriminator (find it — do not assume it is max_steps is not None as in #778), and ENUMERATE the downgrade as a smoke blind spot per .claude/rules/smoke-blind-spots.md.
5. Tests: extend tests/test_datagen_trainability_floor.py (or add a sibling) with the production-raise / smoke-warn / override paths for this entrypoint, and register per scripts/select_step9c_tests.py's WORKFLOW_INVARIANT convention if a new file is added (note the two-place registration — the tuple AND tests/step9c_workflow_invariant_manifest.txt).

Known trap, inherited from #2242

#2242's review logged this Minor on the #778 arm, and #2225 should not repeat it: passing --trainability-floor-override WITHOUT --trainability-override-reason is caught only inside each per-cell subprocess (a ValueError after the tokenizer + dataset load), so an N-cell wave dies N times post-fan-out instead of once at argparse. Add the paired-flag validation in main() before dispatch.

Provenance

Surfaced by the #2242 code review's bug-class sweep (2026-08-12), which verified by grep that scripts/issue2225_train.py calls no gate and that its coupling to issue778_finetune.py is constants-only (zero calls to train_single_cell). Parent #2242 carries the floor's definition, its literature grounding (~12 optimizer steps, the narrowest observed install transition — #533/#547), and the two-tier reconciliation against critic-lens-reference.md item 9(i).