Skip to content

ADR-0080-the-build-overhead-trigger-fired-and-0014-governs-the-gate.md

Latest commit

 

History

History
124 lines (102 loc) · 7.58 KB

ADR-0080-the-build-overhead-trigger-fired-and-0014-governs-the-gate.md

File metadata and controls

124 lines (102 loc) · 7.58 KB
Raw

ADR-0080: The build-overhead trigger fired, the budget was re-expressed rather than the bet withdrawn, and ADR-0014 §4(4) governs the gate

Status: Accepted Date: 2026-08-15 Deciders: Repository owner · Platform architecture

This record reconciles two Accepted records. It decides nothing new and re-argues nothing — every figure below is cited to the record or the benchmark that produced it. It exists for two reasons. A reader following PLAN §1's kill criterion — "Revisit ADR-0002 first" — arrives at ADR-0002 and finds a fired trigger with no verdict attached to it. And 0002 and ADR-0014 disagree in plain text about whether one of 0002's mandatory mitigations binds, with nothing saying which record gives way.

1. The trigger fired

ADR-0002's exit condition is *"> 3 generator defects per delivery phase, or build overhead

8 % sustained"*.

Build overhead at 200 flows — the size P1's exit criterion names — is +67.1 %, 95 % CI [+61.9, +73.6], verdict FAIL, recorded at WP-43 in B12-scale §5.4. Sustained rests on more than that one reading: §5.1 recorded +77.1 % [+72.0, +80.6] on different hardware before the duplicated-bind fix, and the pre-catalogue tree measured +18.4 % [+16.3, +19.9] with the feature that dominates the figure nowhere in the build. Every campaign that has measured this clause has found it true.

The clause's other half — more than three generator defects per delivery phase — is counted by nothing, so it can neither fire nor be shown not to have. Nothing below depends on it.

2. The forced revisit happened, and its outcome was re-expression

The revisit the trigger obliged is ADR-0014, decided 2026-08-10. It kept compile-time orchestration and the derived error catalogue, and changed what the budget says instead.

Why re-expression and not withdrawal. The number was never evidence against the architecture, and 0014 is where that is argued rather than here:

  • Deleting the feature that dominates the figure moves 200 flows from roughly +67 % back to roughly +18 % — still a fail — and no option 0014 priced reaches +8 % (§2).
  • A ratio cannot be passed or failed in the first place. §3 A: a per-unit budget "is testable at any size; a ratio is not, because its denominator is the user's code." The gate's own baseline states it as a measurement — the same generator reads +0.4 % on the one-flow reference sample and +67.1 % at 200 flows, so "a ratio whose bottom half moves with the subject cannot be passed or failed, only re-argued" (generator-cost-baseline.json).
  • The architectural question ADR-0002 actually asked is answered separately, and favourably: the generator is linear in flows — flows^0.95, CI [0.87, 1.06], with no meaningful fixed term, on the same WP-43 run — which is an optimisation backlog rather than an architectural defect.

What binds in its place, per 0014's decision as corrected on the day it was taken (the unit is bytes allocated, not the milliseconds the record recommended, because wall clock moved 139 % across twelve identical runs where allocation moved 0.069 %): the generator allocates ≤ 800,000 bytes per flow and ≤ 160,000 bytes per capability, at any subject size, set from 772,522 / 769,272 B per flow and 148,562 / 146,808 B per capability. Milliseconds stay reported and advisory. That criterion is P1's Done when in 20-Roadmap, and check-generator-cost.py fails the run on a breach.

So ADR-0002's decision stands unwithdrawn and this record does not reopen it. What the trigger obliged was a review; the review happened; this is where its outcome is written down.

3. The gate contradiction, resolved

ADR-0002's first negative declares its mitigation list "all mandatory", and the fourth item on that list is "gate build overhead at ≤ 8 % (budget B12)". ADR-0014 §4(4) commits the opposite: the scale-overhead job "stays advisory until a pass is recorded", and .github/workflows/performance.yml carries continue-on-error: true on it.

ADR-0014 §4(4) governs. Concretely:

  1. The ≤ 8 % clause of ADR-0002's mitigation list is superseded by §2's allocation ceilings — which are absolute, gateable on a shared runner, and blocking. The mitigation is not withdrawn; what changed is the unit it is stated in, and therefore which job enforces it. "All mandatory" continues to mean what it says about a gate that can fail.
  2. scale-overhead stays advisory, and stays red. It measures P1's wall-clock question, which is real and unmet, and generator-cost-gate §7 keeps it visible by reprinting the criterion on every run, including passing ones.
  3. The other three mitigations stand, individually and unchanged — emit readable C# to obj/generated; snapshot-test every emitted file; keep the generator's logic in a pure, unit-testable model layer separate from Roslyn plumbing. Superseding the fourth supersedes only the fourth.

One date does the work, and a reader needs it. ADR-0002's note says the resolving record "is still Proposed" and that the disagreement "stands and is no longer queued for resolution". Both were true when written and stopped being true on 2026-08-10. ADR-0002 is not edited, because this repository does not edit accepted records — so this is where the reader learns it, and the index row is amended to point here.

Consequences

  • A reader sent to ADR-0002 by PLAN §1's kill criterion reaches a verdict in one hop, from the index row.
  • Exactly one record now answers does the ≤ 8 % gate bind — no — and names the criterion that binds instead.
  • This record adds no gate and moves no ceiling. If §2's ceilings are wrong they are wrong in ADR-0014 and its baseline, which is where they are argued and where a correction belongs.
  • Recorded and deliberately not repaired: 0014's own §10 still closes "this record stays Proposed", which its own header contradicts. It is an accepted record's text; the index row now reads Accepted.

Revisit when: a wall-clock overhead measurement at a stated size records a pass, at which point scale-overhead stops being advisory and ADR-0002's fourth mitigation can be restated in its original unit rather than superseded; or one of ADR-0014's four revisit triggers fires on admissible evidence and re-opens the budget this record says governs; or the generator stops being linear in flows — §5.4's flows^0.95 acquiring a superlinear segment, or a fixed term large enough to matter — because §2's answer to why not withdraw the bet rests on that fit and on nothing else.

Back to: ADR index · ADR-0002 · ADR-0014 · B12 at scale