Bounded computation reuse

Status: private engine infrastructure, first admitted by the VED-011 marriage performance work. This standard governs exact reuse within one owned computation scope. It does not establish a public cache API or cross-request service.

Governing object

A memo retains the immutable result of one deterministic computation for the same exact inputs and the same bound dependencies. Reuse removes duplicate execution; it neither interpolates a result nor substitutes a nearby epoch, smaller interval, different source or different numerical regime.

moira/_bounded_memo.py provides bounded_memo(function, key=..., maxsize=...) and binary64_key. The module depends only on the standard library and has no marriage, reader, calendar, frame or policy dependency. Every wrapper owns its own bounded LRU, creating-thread identity and statistics. No global result store retains finished requests or borrowed resources.

Admission contract

1. Bind the callable to one stable dependency scope. A reader snapshot, source tables, correction policy, numerical regime and other unkeyed dependencies must remain fixed for that scope. A changed dependency requires a new scope or explicit cache_clear; the helper cannot infer scientific equivalence.
2. Define an exact bytes key containing every remaining input. Binary64 fields must preserve signed zero and adjacent representable values. A differential key includes value endpoints, derivative endpoints, center endpoints, radius and centered flag. Coarse/precise mode is a separate input.
3. Reuse only immutable primitives, tuples and recursively immutable frozen dataclasses. Mutable containers, lazy iterators and mutable fields inside frozen vessels are rejected before insertion. Do not mutate frozen vessels through low-level Python escape hatches.
4. Keep resource admission, lifetime leases, source checks, cancellation and required accounting outside the memo. A hit does not replay side effects. The pure-computation wrapper must not become a route around a budget check.
5. Bound entry count and each result's size. maxsize=0 retains no results; positive capacities evict the least recently used entry. This is an entry bound, not a claim of a universal process-memory limit.
6. Use the wrapper on its creating thread. Calls, statistics and invalidation from another thread are rejected, including cache hits. Independent requests on different threads create independent wrappers.
7. Exceptions propagate and are never cached. A failed evaluation must be tried again when requested; it cannot become a successful or permanently missing result through reuse. None may be a legitimate immutable result when the owning computation explicitly gives it that meaning.

cache_info() returns detached hit/miss/capacity/size observations; cache_clear() releases entries and resets those observations. __wrapped__ retains the original callable for scoped differential validation. No runtime configuration, optional dependency or public selector is introduced.

First production use: marriage numerical evidence

The marriage request owns five pure-computation memos:

Owner	Reused object	Capacity
AstronomyEnclosures	TT differential to TDB differential	4,096
AstronomyEnclosures	TDB differential and record layout to all admitted record arguments	8,192
FrameEnclosures	TT differential to frame precession polynomials	4,096
FrameEnclosures	TT differential plus coarse/precise mode to full frame angles	4,096
FrameEnclosures	Angle differential to its ordered cosine/sine pair	4,096

All rounded arithmetic schedules, branch families and source-domain checks are retained. Coarse and precise modes share the same precession polynomials only; their nutation enclosures remain distinct. Frame rotations still perform the same ordered operations with the exact original trigonometric results.

The serving pair() owner still selects the reader/descriptor, checks source identity and seams, and charges its work meter before asking for record arguments. Public evaluation/reader/root counters keep their existing meaning. These new memos remove unmetered duplicate arithmetic, not accounted work.

Reuse by another feature

Another engine owner may use the same helper after declaring its own stable dependencies, exact key, immutable result and finite capacity. A tested example outside marriage wraps the existing julian.tt_to_tdb calculation with an exact binary64 epoch key. That demonstrates independence from marriage; it does not enable persistent caching of arbitrary ephemeris results.

Before another production admission, verify key distinctions, dependency isolation, eviction/recomputation, failure retry, ownership/lifetime, budget placement and exact output equivalence against the uncached computation. Keep that product's independent authority and invariant tests. Regression identity demonstrates preservation; it is not a new external astronomical oracle.

See the marriage reuse validation receipt for measurements, source fingerprints and the executed acceptance corpus.