GH-72: Drop the per-nested-object configuration round trip

[magicsunday]

Fixes #72.

Not the suggested fix, and why

The issue proposes that MappingContext hold a JsonMapperConfiguration directly, so toOptions()/fromContext()/replaceOptions() disappear. That is not expressible:

• The option bag is a string-keyed extension point — a type handler puts its own keys there, and no configuration object knows them. CustomContextOptionTest pins a handler-owned key surviving nesting.
• OPTION_ABORT_ON_ERROR deliberately lives only in the context (mapWithReport() plus strict configuration throws instead of returning a report #65): it distinguishes the two entry points rather than expressing a mapping preference, and mapWithReport() sets it alongside toOptions() output.
• Custom context options are wiped on the first nested object and the caller context is mutated #64 exists precisely because replacing the bag wholesale lost custom keys.

So the two representations stay — and the round trip, which is the actual defect, goes. The configuration is translated into the context once at the boundary; the context is what travels.

What made that possible

Inside the mapper the configuration was read exactly twice — isStrictMode() and shouldIgnoreUnknownProperties() — and the context answers both. So the parameter came out of mapSingleObject() and validateAndNormalize(), out of a closure's use list, and the nested strategy callback collapsed to a single expression.

What review caught

I removed the write half and left the read. map()'s third branch rebuilt a configuration from the context for callers supplying only a context — which is every nested object — and nothing consumed it. Eight accessor calls and an allocation per nested object, discarded on the next line. The acceptance criterion "no per-nested-object round trip" was half met.

fromContext() now has no caller in src/. It stays — published API, and a custom strategy receives a context and nothing else — but that is now a stated decision at the method, along with what it does not carry: the mapper's own options only, so a handler-owned key does not survive the trip. Which is why the mapper stopped making it.

The new tests characterise rather than regress. They pass on the revision before this change, as they must for a behaviour-preserving removal. The docblock says so, and says what does make one fail: restoring the round trip together with the replace-instead-of-merge semantics it was written against.

Behaviour

Unchanged, and checked rather than assumed. Restoring the round trip alone leaves the suite green — it is harmless only because #64 made the write a merge. Restore both and four tests fail.

One observable improvement, verified by probe: a caller-supplied context no longer has the eight mapper keys materialised into its raw bag as a side effect of the first nested object. Every accessor reads identically; the bag is just no longer polluted. That matches the property UnionResolutionTest::itDoesNotMaterialiseAnOptionTheCallerNeverSet already pins elsewhere.

Acceptance criteria

• ✅ No per-nested-object round trip — both halves.
• ✅ OPTION_* constants untouched.
• ✅ Behaviour unchanged.

composer ci:test green, verified by exit code: 375 tests, 1216 assertions; PHPStan max, Rector, CGL and CPD clean.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!