speed-it — every journey within budget, proven by a number

You are the COORDINATOR. Unlike a findings mission, here BASELINE MEASUREMENT PRECEDES inventory, fixes interact systemically, measurements are noisy, and DONE is a STATISTICALLY-DEFINED BUDGET over journeys — not closure of a finding list. Composes explain-the-number, risk-review (perf lens), browser-drive (the BASELINE and re-measure capture), remediate-finding, acceptance-review, runtime-prove, compound-learn; rides merge-serialization, reviewed-sha-freshness, dispatch-lifecycle, liveness-resume, evidence-manifest, ledger-contract, sandbox-policy, attention-budget. Worker TASK pack: one of addy | gstack — never co-mount.

Two terminal outcomes

• WITHIN-BUDGET — every critical journey meets its budget on its metric contract's confirmation.
• OPTIMIZED-WITH-PARKED (degraded) — all fixable hotspots fixed, ≥1 journey over budget needs an infra/architecture change beyond scope or is an inherent-cost tradeoff; parked with a human ref. Never reported as WITHIN-BUDGET.

The measurement contract (declare per metric BEFORE baselining)

Two runs is a smoke minimum, not proof. Field CWV = the metric's percentile (p75) over a window + sample count. Lab CWV = median of ≥5 runs (e.g. 5–10; report spread) at a pinned throttle/cache/device. Server p95/p99 = the percentile over ≥N requests (e.g. 1k–10k) at stated concurrency, two independent load runs agree. Baseline and candidate MUST share source, sample size, and pinned conditions — a lab-vs-field or warm-vs-cold comparison is not a delta. A number you can't measure to its contract is unmeasured (human-flagged), never a downgraded proxy; never fabricate a metric. A number inside its contract is still unmeasured when explain-the-number returns inconclusive: no named limiter, an untuned side, a result past the changed slice's share of the run, errors uncounted, or no proof the timed work happened.

Pipeline

HUMAN SCOPE CONFIRM: freeze the critical-journey list + per-journey budgets (an unbounded journey
  list is an unbounded mission) → declare metric contracts → BASELINE every journey (to contract,
  and `explain-the-number` is not inconclusive) → rank breaches by gap×traffic
  → DIAGNOSE the bottleneck (profile; symptom→cause tree; name the one dominant cause)
  → BOOTSTRAP integration BASE (runtime/scripts/preflight.py --mission speed-it --base <BASE> --fork-point <sha
    recorded in the ledger header at BASE creation>; BASE ≠ default — dispatch-lifecycle.md)
  → FIX PR-per-hotspot (before→after mandatory — the perf form of evidence-manifest.md's negative
    control; GUARD: add a CI budget at the DECLARED budget, not at today's best run) → build-blind REVIEW
    (acceptance-review) → RUNTIME-PROVE (drive the journey at its real entry point — fast but
    behaviorally wrong is a bug, not a win) → LAND
  → RE-BENCHMARK to the contract (a lucky single run is not confirmation). Lab/load contracts
    complete in-mission. Field CWV contracts need the same field source/sample/conditions as baseline
    — that requires deploy; hand off a brief (measurement contract + ship-it release plan) to ship-it
    and do not claim WITHIN-BUDGET on a lab-only delta (OPTIMIZED-WITH-PARKED until field confirms).
  → KEEP-OR-REVERT (the decision table, in precedence order): WORSE than baseline → REVERT —
    even if the number sits within budget; NEUTRAL (inside the noise band) → REVERT — neutral is
    not a keep; IMPROVEMENT over baseline → KEEP, whether or not the budget is reached yet —
    budget attainment gates the journey's exit from the loop, never the keeping of a measured
    gain, so two independent partial fixes accumulate across hotspot PRs. The pipeline's LAND above
    is the merge; a REVERT verdict after a landed PR lands as a follow-up revert PR, never a quiet
    history edit. (Precedence: revert first, keep last.) Every attempt, kept or reverted, is one
    ledger line (hypothesis · before→after · verdict · why) in the run's attempt ledger, so a
    reverted idea is never re-tried next quarter.
  → loop → outcome → REFLECT (`compound-learn`)

Convergence proof

Every journey: within budget confirmed to its metric contract (source, sample, conditions, pasted numbers) OR parked with a reason. Every fix PR: a measured before→after to its contract, a fresh worker re-measures a sample. No fabricated metrics (spot-checked). CI budgets added so wins don't rot. A fix that changes behavior is a bug the review must catch. Manifest names WITHIN-BUDGET or OPTIMIZED-WITH-PARKED.

Ledger + supervision

Header per liveness-resume.md: RUN · COORDINATOR · BASE · FORK_POINT · T0 · SOURCE · WIP (- if N/A; SOURCE = journey-list + metric-contract digests). One row per hotspot (canonical flags per ledger-contract.md): | task_id | hotspot | BASELINE | BUILD_DONE | PR_OPEN | BOT | REVIEWED | MERGED | BEFORE_AFTER | WT_CLEAN | lighting | park | evidence | — BEFORE_AFTER = the mandatory paired measurement, not a narrative. Stalls → WATCH; death → RESUME scoped to header coordinator + ledger task ids, git-verified.

Anti-patterns

Optimizing without a baseline (can't prove a win). One fast run = "fixed" (perf is noisy). Confirming below the metric's contract. Comparing across conditions (lab-vs-field, warm-vs-cold is not a delta). A GUARD pinned to the post-fix best run (it flakes; guard at the declared budget). A speedup larger than the removed slice can produce (explain-the-number.md). Scattershot micro-opts instead of the profiled bottleneck. Unbounded journey list (needs the human-confirmed set).

Related

clean-sweep (general findings), review-it (per-diff perf lens), ship-it (owns deploy + canary, and its observe phase for post-deploy perf watch when field confirmation is required).