plan: print every placement term, and a switch to A/B the slab shrink (#885)

[JustVugg]

@mohamedmastouri2000-boop's follow-up on #885 says #869 was half a fix. The measurement that established that took fourteen instrumented runs on a 128 GB host to recover numbers this engine already had in registers.

g64  cap raised 8->31   (v1.4.0: 31, v1.5.0: 17)   the width fix landed
E8   cap raised 8->45   (v1.4.0: 46, v1.5.0: 17)

g64  RSS  105.8 -> 64.1 -> 84.67 GB               and stopped there
E8   RSS  106.5 -> 47.5 -> 75.6  GB

This PR does not attempt the remaining fix. It makes the remaining fix decidable.

The line only printed when the clamp bit

if(pin_bytes+1.0<planned_pin)
    fprintf(stderr,"[PIN] auto: %.1f GB plan capped to %.1f GB ...")

So "the reserve was zero" and "the reserve did not bind" were indistinguishable from outside — and those are the two hypotheses anyone debugging placement is choosing between. It now prints every term, every time:

[PIN] auto: budget X GB | plan Y GB (conf C, N selections) |
      LRU reserve Z GB (cap K/layer x W MB/row-set) |
      max_pin M GB -> pinning P GB  [CAPPED by the LRU reserve]

A reserve near the budget means the pin is starved by a reservation the LRU may never claim — it is demand-filled and only allocates on a miss. That is the open half of #885, and it is now visible at a glance instead of by inference.

@mohamedmastouri2000-boop asked for exactly this, twice, before anyone acted on it. That is the actual lesson here.

COLI_SLAB_SHRINK=0

#869 introduced the only new allocator work in that change and stated the cost was unmeasured. #885 then measured it, on the arm that has it:

v1.4.0 g64   1.21, 1.23              spread 0.02
v1.5.0 g64   1.00, 1.01              spread 0.01
dev    E8    1.43, 1.43              spread 0.00
dev    g64   1.12 .98 1.11 .97 1.05  spread 0.15, bimodal

Every arm on that box replicates to ≤0.02 except the one that is new. Reallocating 19–38 MB per crossed row is an mmap/munmap plus page faults on first touch — a per-run effect, which is what bimodal looks like.

The switch restores grow-only so the hypothesis can be A/B'd without a rebuild. It re-introduces the slab migration, so the per-row cap becomes optimistic again: a diagnostic, not a supported configuration, and it says so on stderr when set.

A duplicate symbol, from a merge of mine

91:  static int g_metal_enabled;   // #587, unconditional, with a comment
                                  // explaining why it must be outside the ifdef
100: static int g_metal_enabled;   // the original, still inside #ifdef COLI_METAL

Two tentative definitions the linker merges — no warning, no symptom, green CI. I verified #587's colibri.c changes were all gated behind g_metal_enabled, which was true, and did not check whether the symbol was already there. Removed; the unconditional one stays.

Tests

test_cap_mixed_width gains case E, which exercises the switch by setting g_slab_shrink directly. It has to: the variable is read from the environment in main(), which the test does not run — so exporting COLI_SLAB_SHRINK would have proved nothing, and on the first attempt it did exactly that, passing identically with the switch on and off.

E. shrink OFF: 892928 B -> 892928 B (must not move)
E. shrink ON : 892928 B -> 450560 B (must come down)

make check: 388 tests, 0 failures.

Refs #885, #869, #815, #587

🤖 Generated with Claude Code