CI: fail on a RAM regression, not only on running out of RAM

[rtissera]

The per-codec RAM ceilings in the three check_budget.py scripts were sized as platform budgets — what the target can afford — so they sit two to eight times above what the code actually uses. That catches a catastrophe and nothing smaller.

Concretely: cd_cdlz used 99014 bytes before the in-place CD spread (#183) and uses 57905 now. The whole 41% win could be handed back tomorrow and the job would stay green, because 99014 is still under the 200000 ceiling.

Two thresholds, because there are two questions

REGRESSION sits about 10% over what the code uses today and answers did we just lose ground. PLATFORM stays exactly where it was and answers does it still fit. Both have to hold, and the output names which one failed:

ok             cd_cdlz    peak=  57905 regression=  64000 platform= 200000
REGRESSION     cd_cdlz    peak=  99014 regression=  64000 platform= 200000
OVER PLATFORM  cd_cdzs    peak= 350000 regression= 142000 platform= 320000

Where the 10% comes from

All seven codecs measured on all three targets, today, on master:

codec	rv32imafc	cortex-m33	hazard3	regression	platform (rv32)
hd_zlib	19766	19766	19766	22000	100000
hd_zstd	107205	107197	107205	118000	150000
hd_lzma	27858	27858	27858	31000	80000
hd_huff	23219	23219	23219	26000	200000
cd_cdzl	49786	49786	49786	55000	200000
cd_cdzs	128795	128787	128795	142000	320000
cd_cdlz	57905	57905	57905	64000	200000

Three different ISAs agree exactly on five codecs and differ by 8 bytes on the other two, so peak heap here is a property of the allocation sizes rather than of the ABI. 10% is therefore generous headroom for allocator and toolchain drift, not a fudge factor.

It is not headroom for design changes. Raising one should happen in the commit that spends the memory, with a reason.

Checked

Against the real qemu logs from all three targets (all pass), against a log edited to put cd_cdlz back at 99014 (fails with REGRESSION where it previously passed), against an over-platform value, and against a missing codec.