json: numbers round-trip exactly — one number→text rule everywhere (#875)

[InauguralPhysicist]

The contract promises num of (str of x) == x. That held for str of and for nothing else.

json_encode, json_build and json_path each carried their own %.15g — one digit short of the 17 a double can need — so a value written as JSON and read back was a different number, silently, in the primary serialization format.

str of      : 3.141592653589793
json_encode : 3.14159265358979        <- decodes to a different double

Worse than filed: integers were corrupted too

Each site also had its own integer fast path, each with a different bound — 2^31, 1e9, 1e9 — so every integer between that bound and 2^53 went through %.15g as well:

1234567890123456  ->  "1.23456789012346e+15"  ->  1234567890123460

An ID silently becomes a different ID. The issue cited only the float case.

The fix

There were four hand-copies of the number→text rule — the fourth in the SIGUSR1 observer dump (vm.c). There is now one, eigs_num_text, and value_to_string calls it too. So json_encode of x == str of x for every number by construction, and a fifth copy cannot appear with a fifth rule.

One deliberate output change

An exact integer below 2^53 now renders bare rather than in exponent form:

json_build of ["v", 1.23e15]   ->  {"v": 1230000000000000}     (was {"v": 1.23e+15})

That is what str of 1.23e15 already produced — the two disagreeing is exactly the drift this removes. At 16 characters it is still nothing like the 22-char fixed-point blob #725 was protecting against, and exponent form is retained above 2^53 (1.23e300 → 1.23e+300), pinned in the same test so the change is bounded. JH97's golden was pinned to the old inconsistency and is updated.

Verification

tests/test_json_roundtrip.eigs pins all three encoders against str of across the hard doubles and the exact-integer band.

Both halves are planted-fault validated, and they turn out to be independently load-bearing:

• capping the escalation at %.15g → the round-trip checks fail;
• narrowing the integer bound to 2^31 → the exponent-form checks fail.

Worth stating: the first plant alone did not catch the second fault. The ragged big integers (1234567890123456) round-trip through the precision escalation whether or not the integer branch exists, so I added the round magnitudes (1000000000000000) that genuinely depend on that bound. Without that, this PR would have shipped a test that looked like it covered the integer path and didn't.

• Release suite: 3787/3787
• ASan + UBSan, detect_leaks=1: 3785/3785, leak tally 0

Closes #875

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Unifies EigenScript’s number→text formatting across str of and all JSON-related producers to guarantee lossless numeric round-tripping (json_decode of (json_encode of x) == x) and eliminate inconsistent integer fast-path bounds.

Changes:

• Introduces eigs_num_text() as the single shared number formatting routine (15→17 digit escalation + exact-integer rendering up to 2^53).
• Switches json_encode, json_build, json_path, str of (via value_to_string), and the SIGUSR1 observer dump to use eigs_num_text().
• Adds/updates tests and documentation to pin JSON encoders against str of, including the exact-integer band behavior change.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/test_json_roundtrip.eigs	Adds round-trip and “agrees with str of” assertions for JSON encode/build/path over hard doubles and exact-integer cases.
tests/test_json_hard.eigs	Updates golden expectation for json_build rendering of an exact integer below 2^53; pins agreement with str of.
src/vm.c	Removes the local number formatting copy from SIGUSR1 observer dump and delegates to eigs_num_text().
src/eigenscript.h	Exposes eigs_num_text() to ensure all producers share one formatting rule.
src/eigenscript.c	Implements eigs_num_text() and routes value_to_string’s numeric formatting through it.
src/builtins.c	Replaces JSON numeric formatting in encode/build/path with eigs_num_text() to ensure exact round-trips.
docs/LANGUAGE_CONTRACT.md	Updates the contract to state that all number-text producers share the same implementation.
CHANGELOG.md	Documents the behavioral fix and the bounded output-format change for exact integers below 2^53.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.