[helly25]

File-diff actions now reuse their resolved normalization settings instead of compiling the same ignored-line regex for every file pair. Targeted native workloads improve about 19% on Linux and 6% on macOS, with existing output and conditional-read behavior preserved.

AG;DR

Base: main after merged #1053 and #1054. Synchronization incorporates current main without changing the verified implementation tree (checked against bc4294e). Use the resulting pushed-head PR-context checks for readiness.

Implementation:

• Prepare one complete immutable DiffOptions record for each distinct action style, after existing whole-command validation. Resolve global defaults, per-action format/context overrides, algorithm, normalization tokens and RE2 ignored-line expressions once per style.
• Pass a const observer through EvalContext. Reached diff callbacks borrow the matching style record after their existing reads, byte-equality and binary branches. The style key remains valid for cloned/deferred expression nodes.
• Preserve lazy context-dependent construction for direct evaluator users and missing table entries. Peer templates/rendering, I/O, patch labels/claims, short-circuiting, safety and output order remain unchanged.
• DiffOptions embeds noncopyable RE2 state: records use ordinary unique ownership, without new reference-counted wrappers, per-file locks, workers, clocks or mutation routes. Comparison-mode patches already have separately prepared options and are unchanged.
• Add an independent in-memory regression target and whole-engine benchmark scenarios. Reconcile O15a with existing EP03/EP07 docs/TODO. Hash-spec/default binding and broader template/executor qualifications remain open.

Measured evidence and limitations:

• docs/benchmarks/prepared-diff-configuration-20261102.md and compressed lossless evidence retain sources, binary/fixture identities, drivers, raw samples, development failures and validation logs.
• Per host: 42 native cases x three sessions; 32 in-memory whole-run cases x three sessions x two variants; six native controls x five confirmation sessions with a byte-identical baseline participant. Discarded warmups, nine rounds/repetitions and fastest-seven estimates remain separate contracts.
• Native normalized 1,000-10,000-file pairs improve about 19% Linux / 6% macOS initially; final confirmation retains about 19-21% Linux / 4-6% macOS. The in-memory 10,000-entry normalized whole-run workload improves about 87% / 88%, including preparation/teardown.
• Preserve costs: final confirmation includes Linux tiny listing +3.59% and macOS tiny listing +1.53%; the earlier prototype's Linux listing +3.67%, tiny plain diff +2.75% and macOS tiny plain diff +1.69% remain separately retained. The report includes duplicate-baseline values and all initial sessions. No universal CLI or first-startup speedup is claimed.
• Unstripped CLI size increases 22,848 bytes on Linux and 1,552 bytes on macOS. Table storage scales with distinct action styles; regex backend bytes are not estimated as table capacity.
• Requested worker counts do not imply concurrent diff actions; those remain ordered coordinator work. Linux native fixtures use tmpfs/affinity; macOS uses warm APFS without affinity. In-memory kernels are distinct from native command timings.

Validation:

• Complete extension-enabled release suites: 321/321 on macOS and Linux, with required Linux FUSE mounts. An earlier lean run with ten skipped targets is retained separately.
• Six focused ASan suites on each host and six focused Linux TSan suites pass. Final regression-test additions receive additional focused checks.
• Direct regressions cover style deduplication/defaults, invalid normalization, no-consumer behavior, prepared/reference output/read counts, skipped reads, missing-style fallback, separate runs and pre-walk errors.
• A genuine compile database was refreshed. All 33 affected translation units pass local lint (32 in the first full pass, then the benchmark source after trailing-comma corrections). A callback-complexity finding was fixed with a pure lookup helper; its final form was rebuilt and remeasured separately. Original prototype data and the first failed hook are retained. Normal enforcing commit/push hooks pass before this PR is created; final readiness depends on current pushed-head CI.