Health-audit fixes + behavior-preserving split of calc_orchestrator God Object

[ackness]

Summary

Two phases from a full health-audit of the workspace, both verified behavior-preserving against the parity gate.

Phase 1 — Audit fixes (correctness / robustness / docs)

• fix(parser): form value extractor now rejects non-numeric captures (FormReject::Nil → text becomes Unsupported) instead of silently injecting a value=0 Modifier into the ModDb. Zero behavior change for inputs that parse today.
• refactor(env_finalize): forward_enemy_modifiers dedup no longer keys on format!("{m:?}") (Debug-format coupling + per-modifier String alloc on a hot path); uses a local field key (f64 via to_bits) instead. Also documents why the charm-limit drop order is already deterministic (single shared ModName bucket, insertion order).
• fix(build-import): surface silently-skipped config inputs / items / text fragments in the PoB2 XML importer (separate Eof from Err, count skipped items, keep lossy text on unescape failure). Import tolerance preserved (parity still 0 parse/calc failed).
• fix(calc-cache): include OrchestratorOptions (base_input + extra_modifier_texts) in the cache key — same Build with different options previously returned a stale cached OutputTable. Adds a regression test; documents that BuildSnapshot::content_hash is not a sufficient calculate_with_data cache key.
• docs: correct workspace/crate/command drift in CLAUDE.md + architecture docs (15 members not 14; pobr-item is implemented; pobr-trade is a standalone leaf; parity test target is parity not ninja_parity; pobr-data-adapter default is warn-and-degrade; deprecation banner on the stale 11-implementation-progress.md).
• test(golden): rename misleading *_GOLDEN_* constants to *_NEUTRAL_* and clarify the suite is a pipeline-smoke + determinism guard (loads no items → all-neutral values), not PoB2 value parity. Assertions/tolerances/values unchanged.

Note: the audit's "charm nondeterminism" finding turned out to be a misdiagnosis — the verify gate's nextest caught it; reverted, with a comment pinning the insertion-order semantics.

Phase 2 — Split the calc_orchestrator.rs God Object (behavior-preserving)

Decomposed the 8782-line calc_orchestrator.rs into a directory module (calc_orchestrator/mod.rs + 9 submodules) via pure code moves + visibility adjustments, zero logic change:

submodule	lines	responsibility
skill_resolve	1000	main-skill / gem property·quality·level / Kalandra / minions
triggers	928	trigger fixpoint + support judgement (thread-local #1)
stat_map	614	StatMap/curse/debuff translation + STAT_MAP_CTX (thread-local #2)
collect	576	passive nodes / jewel expansion / item·gem collection
buffs	504	herald / aura / buff specs / spirit
skill_mods	452	skill base mods / DoT / corpse / crossbow
weapon	428	weapon contribution / unarmed / local parsing
conditions	306	skill_types → flags/conditions
defence	302	armour / evasion / ES / ward

• mod.rs retains the 994-line calculate_with_data spine (intentionally not split — that changes data flow, deferred to a separate task), the public entry points, pub types, and the catch-all test block.
• lib.rs has zero diff (pub mod/pub use is transparent to .rs → /mod.rs).
• Extracted sequentially, one commit per submodule, each independently compiling → clean git bisect.

Verification

• cargo clippy --workspace --all-targets -- -D warnings — clean
• cargo nextest run --workspace — 2092 passed / 0 failed (5 skipped)
• cargo test -p pobr-build --test parity — def25 84.2% / defcore 91.7% / off 76.2% / dot 54.1% @5%, 0 parse/calc failed — zero drift vs baseline (the proof that the refactor is behavior-preserving)

Deferred (separate tasks, each with a reason)

• 994-line calculate_with_data spine split (changes data flow, not a pure move)
• legacy.rs retirement (still the parser fallback path)
• offensive-side parity tightening (panel-mode hit rate is only 48.8%)
• restore the parser-rules drift gate (needs vendor a82a33b bump + golden re-baseline)