fix(homeostatic): the fold UPDATE runs in each backend's own dialect (#736)

[cdeust]

Fixes #736

Cause

_apply_fold (handlers/consolidation/homeostatic_apply.py) issued LEAST/GREATEST in handler code. SQLite has neither, so every fold cycle raised sqlite3.OperationalError: no such function: LEAST on the SQLite backend.

Fix

The UPDATE moves into fold_heat_base(domain, write_class, factor) on each store, as the other heat writers already are. PostgreSQL keeps the same statement; SQLite uses scalar MIN/MAX and CURRENT_TIMESTAMP. No skip, no fallback. The I2 allow-list tracks the two new writer sites.

Evidence

• tests_py/infrastructure/test_sqlite_homeostatic_fold_736.py runs the fold on a real SqliteMemoryStore: fails on main with no such function: LEAST, passes here.
• Homeostatic, I2 and SQLite parity tests pass; ruff check, ruff format --check and the craftsmanship gate are clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VH9JM8qvKiJysP16BmAxAU

[cdeust]

ZETETIC-REVIEW: APPROVE
head: 08be5d9

MIN/MAX scalaires équivalents à LEAST/GREATEST (heat_base est REAL NOT NULL dans les deux schémas). CURRENT_TIMESTAMP suit bump_heat_raw. Le SQL PG est déplacé sans changement. L'allow-list I2 reste exacte, sans relâchement de l'invariant. Couches améliorées : le handler ne fait plus de SQL brut.
Nits non bloquants : le docstring de test_I2 cite encore _apply_fold ; COALESCE(is_protected/is_stale, 0) replie une ligne NULL explicite, alors que PG l'exclut ; no_decay n'est pas couvert par le nouveau test.
Merge après CI verte.

[cdeust]

ZETETIC-REVIEW: APPROVE
head: d3a0743

Delta depuis 08be5d9 (relu par range-diff) : rebase sans changement, plus d3a0743. Le stub acquire_batch sous TYPE_CHECKING correspond à pg_store_schema.py:115. Le test de stratification stub maintenant fold_heat_base et vérifie un seul repli, classe AUTO, jamais DELIBERATE. CI entièrement verte.