fix(mir): close eager continuation lowering

[sanohiro]

Summary

• close the L2b-a2-ac required-child continuation protocol across eager MIR lowering
• maintain O(1) block reachability and reject zero-predecessor joins without emitting parent actions
• preserve temporary ownership and cleanup across terminating children
• add whole-program and per-unit regressions for expression families, source-compatible function values, map_err, and owned-string field views
• split exhaustive handcrafted-HIR validation into the independently mergeable L2b-a2-am-g and L2b-a2-am-b follow-up contracts

Verification

• scripts/pre-pr.sh (includes scripts/test-pr.sh and workspace/all-target Clippy with warnings denied)
• cargo test -p align_mir
• cargo test -p align_driver --test mir_continuation
• cargo test -p align_driver --test expr_depth within_limit_chain_compiles_and_runs
• cargo test -p align_driver --test per_unit_codegen eager_expression_termination
• focused source-compatible callback and owned-string field owners
• clean HEAD/base-bound independent adversarial review

The full expr_depth binary still reproduces its unchanged sema-time deep_within_limit_expression_is_accepted stack overflow on origin/main under local Rust 1.96.1. The MIR-owned depth target above passes, and the debug lower_expr frame improves from 77,168 to 75,808 bytes.

Benchmark

Apple Silicon provenance fixture:

summary-inference          3.136 ms/check   22,848 bytes
import-validation         1.835 ms/import
mir-continuation-lowering 1.105 ms/lower    2,561 blocks

Review closure

The post-open review cycle inspected the complete pushed diff at d2fa63425dc1a5e8754c8c93ddd168eee3ae22d6. Its one valid finding was fixed in the final follow-up: malformed dynamic/SoA field metadata is now rejected before SliceLen or any field action. The malformed-metadata owner test, full align_mir tests, MIR-continuation driver tests, string-view and SoA owners, and align_mir Clippy pass. Per repository policy and explicit user direction, this bounded P2 fix was closed by the finding-to-fix ledger; no second full-diff review was run.