[lua] Implement Sys.getEnv and Timer.stamp for lua-vanilla mode

[jdonaldson]

Summary

• Sys.time() — use os.time() instead of throwing. Note: integer-second precision only (vs microsecond precision with luv). Open question: should we emit a compiler warning here?
• Sys.getEnv() — use os.getenv() instead of throwing
• Timer.stamp() — use os.clock() directly on all Lua targets (subsecond precision, avoids the Sys.time() path entirely)

These are the remaining Haxe-side changes needed for #12843 (hxcoro with -D lua-vanilla). The bit ops part was already resolved by #12599.

Open question

os.time() only returns integer seconds. Should we warn at compile time when Sys.time() is used in lua-vanilla mode, since users might expect subsecond precision? Or is it fine as-is since the docs don't guarantee precision?

Test plan

• CI passes on all targets
• Verify Timer.stamp() returns subsecond values on Lua
• Verify Sys.getEnv() works in lua-vanilla mode

[tobil4sk]

os.time() only returns integer seconds. Should we warn at compile time when Sys.time() is used in lua-vanilla mode, since users might expect subsecond precision? Or is it fine as-is since the docs don't guarantee precision?

I think we should use os.clock for both, unless we specify a behaviour that requires otherwise. That way haxe.Timer.stamp() can continue to call Sys.time().

EDIT: Lua docs say:

The os.clock function returns the number of seconds of CPU time for the program

We already use it for Sys.cpuTime(). So I guess the question is whether it is wrong for Sys.time() to give cpu time.

[jdonaldson]

I looked at what every other target does for Sys.time():

Target	Underlying call	Type	Precision
C++	QueryPerformanceCounter / mach_absolute_time / CLOCK_MONOTONIC	Monotonic	Nanosecond
JVM	System.currentTimeMillis() / 1000.0	Wall-clock (epoch)	Millisecond
Python	time.time()	Wall-clock (epoch)	Microsecond
PHP	microtime(true)	Wall-clock (epoch)	Microsecond
Neko	sys_time native	Wall-clock	Float
HL	sys_time native	Wall-clock	Float
Lua (luv)	luv.Misc.gettimeofday()	Wall-clock (epoch)	Microsecond
Lua (vanilla)	os.time()	Wall-clock (epoch)	Integer seconds

Every target returns wall-clock time (C++ is the outlier with monotonic). Using os.clock() for Sys.time() would make lua-vanilla the only target returning CPU time — which is semantically different. A sleeping/waiting program would show 0 elapsed time with os.clock(), while all other targets would show the actual wall time elapsed.

The precision is unfortunate, but vanilla Lua simply doesn't have a subsecond wall-clock API without external libraries. os.time() is the correct semantic match — it just has integer-second granularity.

For Timer.stamp() I think os.clock() is fine since it's already used for Sys.cpuTime() and the docs describe it as a monotonic timestamp for benchmarking.

[jdonaldson]

I remember not using os.time() for this reason. I think it might be ok for lua-vanilla, maybe with some sort of warning? I wasn't using it for anything critical, so I'm all ears on what the right tradeoff is here.

[jdonaldson]

Friendly ping on this one — wanted to check whether my Mar 27 reply changed your thinking on os.time() vs os.clock().

To recap the tradeoff: os.clock() returns CPU time (a sleeping program shows 0 elapsed), so using it for Sys.time() would make lua-vanilla the only target whose Sys.time() doesn't track wall-clock. os.time() matches every other target's semantics, just at integer-second precision.

Three options I see, happy to go with whichever you prefer:

1. os.time() as-is — correct semantics, integer precision, no warning.
2. os.time() + compile-time warning when Sys.time() is used in lua-vanilla mode, so users aren't surprised by the precision drop.
3. Document the precision caveat in the API docs only, no warning.

Just rebased onto development to clear any merge-ref staleness.

[jdonaldson]

Dropped the Sys.time() implementation from this PR. Vanilla Lua's os.time() only has second-level precision, so it would silently disagree with the fractional-second wall-clock time Sys.time() returns on every other target — more confusing than helpful. I've left it as notImplemented() in lua-vanilla mode; subsecond wall-clock time genuinely requires luv (gettimeofday), which is what non-vanilla mode already uses.

The PR now just implements Sys.getEnv() (via os.getenv()) and Timer.stamp() (via os.clock()), both of which have exact vanilla equivalents. Force-pushed and rebased onto latest development.