The records oracle idles 60 s before every timed das cell; ledger rows 138 and 140 carry their verdicts

[borisbat]

Behavior change: the records oracle idles 60 s before every timed das cell (was 12 s) - an oracle board takes longer per box.

Why. The oracle's first read after a big cell lands 20-33% under its stored mean and spends a FAIL and a retry slot. The previous cell's heat outlives the 12 s reclaim settle: on the M5 Max the GPU still runs 990-1420 MHz instead of 1620 when the next cell starts.

What changes.

• --oracle-settle (default 60) idles before each timed oracle das cell on every leg; the reclaim settle stays 12 s.
• The three sleep-if-positive helpers fold into one settle(seconds); the retry settle goes through it too.
• Ledger row 138 carries the throttle's cause and provenance; row 140 closes the CPU ASR shortfall as core placement on the box, not a code regression.
• Three checklists fix what applying them to this diff surfaced (served-turn and board-cell definitions, figure rules keyed on legs of a turn, diff-scoped records rules, one override boundary, every STYLE037/038 suppression spelling).

Observable behavior.

• oracle cell after a big cell reads 20-33% low and retries -> reads within 1% first time
• --oracle --legs metal on the M5 board -> about 17 min longer (21 cells x 48 s of idle)

Where to look. The three settle(cfg.oracle_settle) sites in performance/gen_bench_records.das and the two rewritten ledger rows.

Validation, claims, ledger

Validation

• Full preflight ran once; compile-sweep was red on a stale binary (the LSP watchdog merge added a fio binding after the last build) and passed on the targeted rerun after an incremental rebuild.
• benchmarks/REVIEW.md applied by hand to the driver: no instrument, clock, or kernel dispatch is added, so its rules do not fire.

Claims - stated, not tested

• The --oracle-settle arm, its three sites, and the folded settle guard have no test: the tool drives model processes for hours and its three sibling knobs carry none either. The measurement stands in: five of five reproductions of the cold read with the board's sequence, 60 s and 120 s settles reading Nominal within 1%. A break would show as an oracle board whose first cell after a big one fails and passes its retry.
• 60 s is the M5 Max's number; the Vulkan boards show the same first-read drop but were not instrumented.

Not done

• Row 138 keeps the thermal-state wait (NSProcessInfo.thermalState) as the unquirk that replaces the fixed number.
• Two lint candidates from the checklist review: a REVIEW.das cell for the records/mtp/ engine-stamp rules, and the defaults/ version-pin compare in check_defaults_profiles.

[Copilot]

🟡 Changes recommended

The new settle helper uses a //! doc comment in a non-attaching position and multiplies seconds * 1000 as int before casting, which can overflow for large values and should be corrected.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Updates dasLLAMA’s performance “records oracle” workflow so oracle runs idle longer before each timed das-cell (default 60s vs 12s) to reduce post-heavy-cell thermal throttling skew, and documents/records the updated measurement rationale in module docs and checklists.

Changes:

• Add --oracle-settle (default 60) and apply it before each timed oracle das cell; keep reclaim settle behavior separate.
• Fold multiple “sleep-if-positive” helpers into a single settle(seconds) helper (also used for retry settle).
• Update dasLLAMA measurement/review documentation and follow-up ledger rows to reflect the revised interpretation and the new settle behavior.

File summaries

File	Description
modules/dasLLAMA/REVIEW.md	Clarifies checklist routing and override/suppression rules language.
modules/dasLLAMA/REVIEW_MEASUREMENT.md	Tightens definitions around served-turn legs, board cells, and provenance requirements.
modules/dasLLAMA/PROFILE.md	Documents the new oracle settle cost and the motivation/observations behind it.
modules/dasLLAMA/performance/REVIEW.md	Refines performance records checklist wording and diff-scoping for records artifacts.
modules/dasLLAMA/performance/gen_bench_records.das	Implements --oracle-settle and consolidates settle helpers used across oracle + retries.
modules/dasLLAMA/followup_general.md	Updates ledger rows (138/140) with revised findings and next-step “unquirk” direction.
modules/dasLLAMA/BRINGUP.md	Updates oracle-mode operational guidance to reflect the new settle behavior.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟢 Approval recommended

The changes are cohesive, low-risk, and the new --oracle-settle behavior is consistently applied at the intended oracle call sites with corresponding documentation updates.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🟡 Changes recommended

The updated LSP test’s publishDiagnostics filtering can match unrelated documents, risking false passes/failures unless it filters by params.uri.

Get a fresh assessment by requesting another Copilot review.

Review details

Suppressed comments (1)

tests/lsp/test_lsp_protocol.das:371

• The stale counter increments on any non-empty publishDiagnostics while waiting for the definition response, but it does not check the diagnostics are for the test document uri. Filtering by params.uri avoids false failures if the server publishes diagnostics for some other open/managed file during this window.

                let def_resp = read_until(r) $(js) {
                    if ((js?["method"] ?? "") == "textDocument/publishDiagnostics" && diagnostics_of(js) != 0) {
                        stale++
                    }
                    return (js?["id"] ?? -1) == 2

• Files reviewed: 8/8 changed files
• Comments generated: 1
• Review effort level: Lite