glossary.md

Performance glossary

Terms the performance skills and techniques.md use, defined for this plugin. A link points at the entry or skill section that applies the term; where a term has an upstream source, that catalog entry carries the context-glue record pointing at it.

• Arithmetic consistency check. Predict the magnitude a proposed cause should produce and compare it to the observed value; a mismatch rules the cause out. See E.
• Baseline window. The time new instrumentation needs to accumulate enough data to give a number. See B.
• Ceiling. The checked-in maximum a ratchet enforces. See ratchet.
• Census. A count of everything that runs on one interaction: subscribers, hooks, observers, process spawns. See A.
• Code cache. Compiled startup code stored so a launch can skip compiling it. See F.
• Cold start, fresh load, warm navigation. Three start states: nothing cached or running; a new page load with the process warm; a navigation inside an already-loaded app. Each is a separate measurement. See B.
• Correlation. Evidence that moving a proxy moves the metric users feel. unproven is a legal value; a proxy with unproven correlation may be measured but not claimed. See D and /performance:goal §1.
• CPU hitch (long task). A main-thread stall long enough for users to notice. See the hitch sweep in E.
• Cumulative layout shift. A composite score summing layout shifts over a page's life. It can stay green while individual shifts are visible. See layout shift.
• Deterministic counter. See drift-immune counter.
• Deterministic stepping. Advancing time in controlled ticks (for a browser, DevTools begin-frame control) so each unit reads the same on every run. See the rig checks in C.
• Diminishing returns. The point where further gains in a workstream shrink below their cost; the cue to close it. See I.
• Dogfooding. Releasing to internal users first, who surface defects no metric sees. See staged rollout.
• Drift-immune counter. A count (instructions, spawns, queries, renders) that does not vary with host load, so it reproduces where wall-clock time does not. Two runs on an unchanged subject must agree before it is trusted; ratchet.py measures twice by default, and --runs tunes it. See /performance:target "Name the counter, not just the duration" and C.
• Feature flag. A runtime switch that gates a change. Here, removed once it is safe, and classified as a kill switch or a ramp. See H.
• Floor. The irreducible cost of an operation, measured on its cheapest possible version before any work. See /performance:goal §2.
• Frame budget. The time one frame may take at a refresh rate: 1000 ms divided by the rate in hertz. See B.
• Geometric mean. The nth root of the product of n values. Use it to combine speedup ratios: a 2x speedup and a 2x slowdown average to 1x, no single large ratio dominates, and the result does not depend on which side is the reference. See J.
• Guardrail. A test, check, or field alert that fails when a proven win regresses. See G.
• Hidden work. Cost that falls outside every headline metric's window, such as a reload after the load metric has ended. See E.
• Hill climbing. Repeatedly changing the code and keeping each change that moves a re-runnable number the right way. It needs a number first. See A.
• Hot path. Code that runs on every instance of a frequent operation (startup, each keystroke, each tool call), where small costs multiply. See A.
• Hydration. A client framework attaching to server- or statically-rendered markup and taking over. The static shell is shown until it finishes. See static shell.
• Journey. One user task measured end to end, such as opening a project or submitting a form. See A.
• Kill switch. A flag whose purpose is to turn a change off if it breaks. See feature flag.
• Lab. Measurement on a controlled rig the agent can re-run, as opposed to real-user monitoring. See C.
• Latent-defect exposure. A speedup changes timing and reveals a defect that was always present. See the diagnosis ladder in E.
• Layout shift. Visible movement of page content. As a defect, exclude movement the user caused and movement before the page is usable. See B.
• Long task. See CPU hitch.
• Main-thread blocking time. Total time the main thread is busy long enough to delay input or rendering. See J.
• Measurement boundary. The start event, end event, and client or server side that define what a duration covers. See B.
• Measurement matrix. Journeys crossed with platforms and products: the full list of measurements to baseline. See A.
• Megamorphic lookup. See polymorphic lookup.
• Memoize finished work. Caching the processed form of content that will not change, so each update costs only the new part. See F.
• Named region and phase. Labels on an event: the UI area it affected and the load stage it occurred in. See C.
• Percentile (p50, p95). The value below which that share of samples falls. The plugin's house default is p50 and p95. See /performance:goal "Percentiles and sample count".
• Placeholder handoff. The moment the real UI replaces a static copy; it must move nothing and lose no input. See F.
• Polymorphic lookup. A property or dictionary lookup whose object shapes vary, which a JIT cannot specialize; megamorphic when the shapes are many. A hot-path cost signal. See F.
• Prerender. See speculative loading.
• Progressive reveal. Showing a large structure incrementally so no single update blocks. See F.
• Proxy metric. A cheaper, steadier number standing in for what users feel. Valid only with proven correlation. See D.
• Ramp. A flag whose purpose is gradual exposure. See feature flag.
• Ratchet. A checked-in ceiling CI enforces: a rise fails the build, a fall lets the ceiling be lowered. See G and /performance:protect.
• Real-user monitoring (field). Measurement from real users' sessions, segmented by build, platform, and environment. See H.
• Realistic and ideal targets. The expected result given the floor and baseline, and the cost with no incidental overhead, held separately. See /performance:goal §3.
• Red N/N and green N/N. A binary check that fails on every one of N runs on the base and passes on every one of N runs on the change, N chosen per check and recorded. See harness-integrity.md rule 3.
• Rig ceiling. A cap the rig imposes on the metric, such as a fixed tick rate, which hides headroom. See C.
• Sidequest. A side investigation allowed to run while it shows signal. See I.
• Speculative loading. A browser fetching or rendering a page before the user navigates to it; the page can be laid out for a viewport other than the one it is shown in. See E.
• Staged rollout. Exposure in stages: internal users, a small share of users, everyone. See H.
• Static shell. A static, usable version of the main UI served in the first HTML response and shown while the framework starts. See F.
• Thread. One narrow workstream with one owner. Mapping it to an agent session is this plugin's inference. See I.
• Thread owner. The one named person who makes a workstream's rulings. See I.
• Time to usable. Time until the user can act, such as the input accepting keystrokes. See B.
• Warm navigation. See cold start, fresh load, warm navigation.
• Win decay. Erosion of a performance win as other changes land. See G.
• Worker offload. Moving CPU work off the main thread into a worker. See F.