docs: state the zero verdicts' resolution; drop the shipped v2.7.x plan row

CodeRabbit round 2 on #555, both taken.

1. docs/performance.md: the five zero verdicts now say what "zero" means
as a number. The order-bias controls moved by under 0.6% on most
workloads and by up to 2.4% on a few (A run 2, C run 2, IMP-06), so each
candidate is read against its own control, and an effect below roughly
half a percent cannot be excluded from these runs. The method's
sensitivity at that scale is shown, not assumed: the same A/B/A,
pointed at a path the workload executes (§3.5b on spritecans.nes),
resolved 0.10% and 0.14% at p <= 0.02 against flat controls. No new
positive-control run was made; the §3.5b measurement already is one on
the same procedure.

Correction inside this commit's own drafting: a first version said the
controls moved "up to about 0.3%". The table in the same section shows
+1.89%, +2.04%, +1.97% and -2.43%; the text now matches the table.

2. VERSION-PLAN.md: "Planned next (not yet released)" still listed v2.7.x,
which has shipped in full (v2.7.0 through v2.7.6). Stale since v2.7.5.
Row removed; v2.8.x is now the first planned line. (CodeRabbit's line
reference pointed at the shipped-history row, which is correct where it
is; the finding's substance was the planned table.)

Verified: release_anchor_audit 15 and release_state_prose_audit 8 pass;
markdownlint pass.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014qfTKi2M3swo7qnwvYCkDj

Author: doublegate

VERSION-PLAN.md

@@ -64,7 +64,6 @@ released version; these are plans, and `to-dos/plans/` holds the detail.
 
 | Version | Scope | Plan |
 |---------|-------|------|
-| v2.7.x | The **core and frontend audits** ([`docs/audits/`](docs/audits/README.md)): untrusted-input hardening and a snapshot fuzz target, battery-save integrity, the mapper gaps, desktop/web/mobile host fixes, then performance only where it measures | [`v2.7.x-core-frontend-audit-plan.md`](to-dos/plans/v2.7.x-core-frontend-audit-plan.md) |
 | v2.8.x | The **libretro and RTL audits**: FFI safety and build fixes in the libretro core; RTL correctness and robustness; the off-die SDRAM build; one seed sweep per build at the line's close | [`v2.8.x-libretro-rtl-audit-plan.md`](to-dos/plans/v2.8.x-libretro-rtl-audit-plan.md) |
 | v2.9.x | **Re-audit, optimise, and the board**: the audits re-run against the fixed tree, final seed sweeps, the release-candidate `.rbf` pair, and the bring-up on the SuperStation One | [`v2.9.x-final-audit-and-hardware-plan.md`](to-dos/plans/v2.9.x-final-audit-and-hardware-plan.md) |
 | v3.0.0 | **The SuperStation One core**: the first hardware-verified FPGA bitstream. On-die `.rbf` as the headline, off-die as a labelled secondary; a MAJOR bump under [ADR 0041](docs/adr/0041-hardware-release-is-v3.0.0.md) | [`v3.0.0-superstation-core-plan.md`](to-dos/plans/v3.0.0-superstation-core-plan.md) |

docs/performance.md

@@ -1034,7 +1034,19 @@ it is not (A's `flowing_palette_fast` −0.26% in run 1, +1.93% in run 2; B's
 `nestest` −0.23%, then +5.46%; C's two exact-path workloads flat, then +1.5% and
 +1.8%).
 Those are the layout and background-load effects the method exists to catch,
-and none of them points toward a gain. **Five are zero, individually.**
+and none of them points toward a gain. **Five are zero, individually** --
+meaning zero within what this method can resolve, which is worth stating as a
+number rather than leaving "zero" to read as "none". On this loaded desktop the
+order-bias controls moved by under 0.6% on most workloads and by up to 2.4% on
+a few (A run 2, C run 2, IMP-06), which is why each candidate is read against
+its own control and never alone. Read that way, a real effect below roughly
+half a percent on these workloads cannot be excluded from these runs. The
+method's sensitivity at that scale is demonstrated rather than assumed:
+the same A/B/A procedure, pointed at a path the workload does execute (§3.5b on
+`spritecans.nes`, below), resolved a 0.10% and a 0.14% ceiling at p <= 0.02
+against flat controls. The five deletions produced nothing of that kind on paths
+the benches demonstrably execute (§3.6's counted, the rest per cycle or per
+dot).
 
 **§3.5b was never reached, here or in v2.7.5.** `Pulse::output` tests
 `length.count == 0` first and short-circuits, and both bench ROMs are silent: