build/flags.stamp invalidation is target-independent: a maintenance target that builds no PRG deletes every object and the PRG, and leaves a stale .map behind

[JC-000]

build/flags.stamp invalidation is target-independent: a maintenance target that builds no PRG still deletes every object and the PRG, and leaves the stale .map behind

Summary

#159's stamp compare runs at Makefile parse time, which is the whole point of the design and is correct. The consequence that is not written down anywhere: it runs whatever target you asked for. make ip65-libs and make ip65-blob — neither of which produces a PRG — will delete every .o and build/c64-https.prg if their (flagless) command line expands differently from the stamped one.

Worse, and this is the part that actually costs you: build/c64-https.map is NOT deleted. You are left with a map file that corresponds to no artifact in the tree, in a repo where every memory-margin figure is measured off exactly that file.

This is not a bug report against #159's behaviour, which is sound and enforced by tools/test_build_flags_stamp.py. It is a request to (a) document the target-independence and (b) decide whether the surviving .map is acceptable.

Measured

All from a clean checkout at master 02e5662, libs/nistcurves v0.14.0.

Baseline — make BACKEND=uci USE_NISTCURVES_ONCHIP_COMB=1: 24 objects, PRG 00e4370ce3086e8b30df243861e56725a09797f08d8e668535c5166c62882152.

invocation after that build	objects	PRG	map
make -n ip65-libs	24	present	present
make ip65-libs	0	gone	present (stale)
make ip65-blob	0	gone	present (stale)
make BACKEND=uci USE_NISTCURVES_ONCHIP_COMB=1 ip65-libs	24	present	present

So:

1. It is not ip65-libs-specific. ip65-blob does it too. It is any real invocation whose expanded CA65FLAGS/LD65FLAGS differ from the stamp — the target is irrelevant, because the compare happens before target selection.
2. Carrying the same flags through preserves everything (row 4). That is the workaround, and it is not obvious: you have to know to repeat build flags on a target that does not use them.
3. -n is correctly exempt and even says so, which is make -n deletes all objects and the PRG: the flag-stamp compare runs at parse time #174 working exactly as designed:

   Makefile:781: A real build would rewrite the stamp and delete every object and build/c64-https.prg
   Makefile:782: at parse time. NOTHING was deleted; the tree is untouched.
   

Why it matters here specifically

CLAUDE.md's Memory-layout section demands that margin figures be re-measured in all five profiles off build/c64-https.map, and names "corrected in one profile, left stale in the other four" as the recurring failure. The natural way to do that is: build a profile, read the map, build the next. If anything in that loop shells out to a maintenance target without repeating the flags, the map silently stops matching the PRG — and a .map that survives its .os gives no signal that it has gone stale. Every number in it still parses fine.

This was hit for real while measuring the uci-comb CRYPTO_OVERLAY tail for #193 (the 153 B → 126 B correction on fix/triage-quickwins); the measurement was saved first, by luck rather than by rule.

Suggested resolution

Pick one or more:

• Docs (cheapest). CLAUDE.md's Build section describes flags.stamp at length but frames it entirely as protecting builds. Add that the compare is target-independent, that a flagless maintenance target therefore invalidates, and that repeating the flags is the way to avoid it. The make ip65-libs / make ip65-blob bullets are the place a reader would look.
• Delete the .map alongside the PRG. One line in the invalidation, and it converts a silently-stale measurement source into an absent one — the same argument $(PRG)'s recipe already makes for rm -fing the target ("an absent PRG is the honest state after a failed link"). The map deserves the same treatment for the same reason.
• Exempt the maintenance targets. Probably wrong — it reintroduces exactly the "make skipped an invalidation it needed" hazard make -n deletes all objects and the PRG: the flag-stamp compare runs at parse time #174 fought, and the exemption logic is already the subtlest part of the Makefile. Recorded so it is visibly considered and declined.

My preference is docs + deleting the map; the third option is noted only to be rejected.

Not claimed

I have not checked whether make package / package-verify are exposed to this, nor whether any test suite or rig shells out to a maintenance target mid-measurement. Both are worth a look and neither is in evidence above.