refactor: simplify accessor internals

[Koan-Bot]

What

Three targeted simplifications to lib/Simple/Accessor.pm addressing issues #22, #25, and #26.

Why

• require mro was called on every new() invocation inside the hot-path _all_attributes() — unnecessary since mro is a core pragma that only needs loading once (optimization: require mro called inside hot-path function on every new() #22)
• String eval qq[require $module] in with() is a code injection risk when a simpler block eval works — the module name is already validated by regex, but block eval is inherently safer (simplification: string eval used for require where a block eval suffices #25)
• The _after_* rollback logic was copy-pasted in two branches of the accessor closure (class hook vs role hook), creating a maintenance trap for future rollback changes (duplication: _after_* rollback logic copy-pasted in two branches of the accessor #26)

How

• Moved require mro to module top-level (load once at use time)
• Converted eval qq[require $module; 1] to eval { require $file; 1 } with :: → / path conversion
• Unified hook dispatch: resolve coderef first (class → role chain fallback), then single call + single rollback block. Same pattern applied to builder resolution.

Testing

Full test suite passes (all existing tests cover these code paths — roles, hooks, builders, transitive composition, inheritance, MRO).

Closes #22, closes #25, closes #26

---

Quality Report

Changes: 1 file changed, 21 insertions(+), 23 deletions(-)

Code scan: clean

Tests: passed (OK)

Branch hygiene: clean

Generated by Kōan post-mission quality pipeline

[Koan-Bot]

Same situation as #27 — this has merge conflicts and the codebase has evolved significantly since this was opened. The incremental fixes (hook cache, strict constructor timing, etc.) have addressed some of the same concerns. Let me know if you'd like me to rebase or if we should close this one.

[Koan-Bot]

Closing: same rationale as #27 — superseded by PR #34 (covers #22 + #25 with tests) and master's hook cache (commit 6b8f5d7 solved #26). This PR also has merge conflicts. PR #34 is the definitive version.