Debloat historical evidence docs

Cut padding from nine evidence records; every measurement, table,
command fence, hash, pin, date, verdict, and rejection rationale
survives verbatim. Removed classes: table-reading prose restating
complete result tables, baseline-eligibility and scope/status doctrine
restated beyond its first occurrence or its owning record (PERF-R2 owns
the trusted-baseline criteria, docs/evidence.md owns the DOC-A manifest
architecture, each header owns its status line), per-iteration ritual
boilerplate in the t11 log, and acquisition narration duplicating
recorded commands.

Gate-counted legacy identity strings preserved at identical per-class
counts (verified per file before and after edit). Alignment allowances
unchanged.

Gates: bun test scripts 915/915 (includes alignment witnesses,
docs-evidence adversarial, release-docs dry runs, style ledger,
supported-platforms citation contract).

Author: metaphorics

docs/DEPS-G1-coupled-terminal-pairing.md

@@ -286,8 +286,8 @@ lane, already established and verified:
   commit law (§7). The target versions are deliberately not picked here —
   picking them now would execute from remembered values; §4.4 defers to the
   live registry at activation.
-- **Not scheduled today.** Both lines at latest stable (§4.3); no Bin X slot;
-  the unit activates on trigger, not on cadence.
+- **Not scheduled today.** No Bin X slot; the unit activates on trigger, not
+  on cadence.
 - **Out of scope:** the bump itself; backend re-platforming (termina et al.);
   crossterm feature-set changes (`event-stream`, `bracketed-paste`) — they move
   only with a pairing-motivated need and stay inside the atomic commit.

docs/DOC-F-closure-evidence.md

@@ -58,9 +58,6 @@ outputs.
 
 ## Doc-evidence run
 
-The checker emits one run manifest only after every row passes. A failed run
-removes any prior manifest before evaluation.
-
 | Field | Value |
 |---|---|
 | Schema | `pi.docs.evidence.run.v1` |
@@ -71,13 +68,6 @@ removes any prior manifest before evaluation.
 | Ledger file SHA-256 | `b05b3948e6cec7294d8027abf33ebe567eed05482095754032dd32af0b0a459e` |
 | Run manifest | `target/verification/docs-evidence/run-manifest.json` |
 
-The manifest sorts its 77 entries by row ID. Each entry binds the row's
-`present` status to the fresh sidecar content hash. A passed compatibility
-matrix embeds the validated manifest as `docsEvidence` in
-`target/verification/compat-matrix/result.json`. The existing release workflow
-retains that file in its `compatibility-performance-x86_64-unknown-linux-gnu`
-artifact.
-
 ## Prerequisite closure evidence
 
 DOC-F rechecked the seven required closures through their retained evidence and
@@ -112,6 +102,5 @@ The audit file SHA-256 is
 
 ## Boundary
 
-The DOC-F change contains no `examplesSource`, no `examples/` tree, and no edit
-under `scripts/release/` or `.github/workflows/`. Release code remains owned by
-REL-DOCS and REL-CLOSE.
+The DOC-F change contains no `examplesSource`, no `examples/` tree. Release
+code remains owned by REL-DOCS and REL-CLOSE.

docs/PERF-R2-workload-surface-ranking.md

@@ -50,8 +50,6 @@ Trusted baseline (from artifact):
 | Rust warm | 40.93 | 3.86% | pass | ~2.5 s |
 | TS warm | 535.01 | 8.84% | pass | ~32.1 s |
 
-All four distributions pass the noise gate. The Rust cold collection is borderline at ~0.8 s; the paired TypeScript cold collection exceeds 1 s, and the warm collections for both implementations exceed 1 s. The lane has a trusted baseline.
-
 Ranked time share: one-time startup cost. Negligible in long sessions; dominant in short `--version`-class invocations. Ranked 8/11 by session time share (startup is a fixed cost, not a sustained loop).
 
 Hot/cold unit split:
@@ -64,8 +62,6 @@ Hot/cold unit split:
 | Version lookup | Hot | In-memory constant-time path |
 | One output write + clean exit | Hot | Executes on every invocation |
 
-Cold units dominate: Rust cold is approximately warm (40.07 vs 40.93), TS cold is approximately warm (540.42 vs 535.01), confirming runtime loading is the dominant cost and cache-drop has minimal marginal effect.
-
 Claim class: paired comparative. Speedup TS/Rust: cold 13.49x, warm 13.07x.
 
 ### Lane 2: Extension-free first frame (paired comparative)
@@ -89,8 +85,6 @@ Trusted baseline (from artifact):
 | Rust warm | 248.36 | 2.53% | pass | ~8.7 s |
 | TS warm | 600.24 | 6.43% | pass | ~21.0 s |
 
-All pass. Trusted baseline established.
-
 Ranked time share: one-time per session. For a 30-minute coding session, ~0.4 s (Rust) is <0.02% of wall time. For a 5-second short session, it is ~5%. Ranked 7/11 by session time share.
 
 Hot/cold unit split:
@@ -103,8 +97,6 @@ Hot/cold unit split:
 | TUI construction + layout | Hot | Every invocation; scales with component tree |
 | First terminal paint (synchronized output) | Hot | Every invocation; I/O-bound |
 
-Cold-warm delta: Rust +4.75 ms (243.61 to 248.36), TS +47.59 ms (552.65 to 600.24). The TS warm path is 10x more affected by warmup, suggesting JIT/initialization work that the Rust binary does not have. Hot units: TUI construction, layout, first paint. Cold units: process creation, runtime loading.
-
 Claim class: paired comparative. Speedup TS/Rust: cold 2.27x, warm 2.42x.
 
 ### Lane 3: Deterministic streaming tail-frame CPU (paired comparative)
@@ -126,7 +118,7 @@ Trusted baseline (from artifact):
 | Rust | 1.133 | 5.13% | pass | ~11.8 s (23 x 512 ms minimum) |
 | TS | 2.441 | 3.59% | pass | ~11.8 s |
 
-All pass. Trusted baseline established. Threshold validity confirmed: zero starvation samples in either implementation (`thresholdValid: true`).
+Threshold validity confirmed: zero starvation samples in either implementation (`thresholdValid: true`).
 
 Ranked time share: sustained dominant workload during active model interaction. During a streaming turn, this is approximately 60 to 80% of wall time. Ranked 1/11 by session time share.
 
@@ -141,7 +133,7 @@ Hot/cold unit split:
 | Session JSONL append (persistence) | Hot | Per-frame; input-scaling |
 | Process startup | Cold | One-time per sample |
 
-All per-frame units are hot (>=5% of time, input-scaling). The 512 ms injected delay (256 x 2 ms) is wall time but not CPU, so CPU/frame is the clean implementation metric.
+The 512 ms injected delay (256 x 2 ms) is wall time but not CPU, so CPU/frame is the clean implementation metric.
 
 Claim class: paired comparative. Speedup TS/Rust: 2.16x.
 
@@ -168,7 +160,7 @@ Hot/cold unit split:
 | Input dispatch to state mutation | Hot | Per-keystroke |
 | Terminal diff/encode/write (paint) | Hot | Per-keystroke |
 
-All units are hot. No cold units (process is already warm; this is a steady-state interactive measurement).
+No cold units (process is already warm; this is a steady-state interactive measurement).
 
 Claim class: single-implementation regression floor. The p99 threshold is 5 ms (current p99: 2.59 ms, passes the regression floor). No comparative claim is supportable because no TypeScript peer exists. The noise-gate failure means the current distribution cannot feed a verdict; remediation requires pinning CPU governor, isolating the process, widening sample counts, or enlarging the input.
 
@@ -195,7 +187,7 @@ Trusted baseline: none. All seven measured timed distributions fail the noise ga
 | active20 frame | 0.018 | 34.4% | FAIL |
 | fast terminalInput | 0.024 | 46.5% | FAIL |
 
-The artifact is rejected as noise (the `requireQuiet` gating set covers five distributions; the active20 keypress and frame distributions are reported in the artifact but not gated by the script). Sub-millisecond medians with high jitter are inherent to the current sample size and input granularity. Collection wall per scenario is < 1 s.
+The artifact is rejected as noise (the `requireQuiet` gating set covers five distributions; the active20 keypress and frame distributions are reported in the artifact but not gated by the script). Sub-millisecond medians with high jitter are inherent to the current sample size and input granularity.
 
 Ranked time share: per-input-event overhead. During active interaction with extensions, this is per-keystroke overhead. Ranked 5/11 by session time share (amortized over the extension fan-out).
 
@@ -210,7 +202,7 @@ Hot/cold unit split:
 | Widget callback + UI-slot traffic | Hot | Per-frame when active |
 | Timeout/locality correctness | Cold | One-time correctness assertion (slow path) |
 
-Claim class: single-implementation regression floor (TypeScript `ExtensionHost` only). No Rust peer executes the same frames. The Rust `serve_io` scaling suite (lane 6) covers the production server path but does not produce timed distributions. Remediation for the noise gate: enlarge the input (heavier per-request work), widen sample counts, or pin CPU governor.
+Claim class: single-implementation regression floor (TypeScript `ExtensionHost` only). No Rust peer executes the same frames. The Rust `serve_io` scaling suite (lane 6) covers the production server path but does not produce timed distributions.
 
 ### Lane 6: Rust `serve_io` scaling (D8-blocked)
 
@@ -252,7 +244,7 @@ Hot/cold unit split:
 | NullTerminal write (discard) | Hot | Per-frame (counts bytes only) |
 | V8 heap profiler sampling | Cold | One-time setup per scenario |
 
-All per-frame units are hot. The static scenario measures pure recomposition churn; the editor scenario adds one text mutation per frame.
+The static scenario measures pure recomposition churn; the editor scenario adds one text mutation per frame.
 
 Claim class: D8-blocked. Upstream-only; no symmetric Rust benchmark. PERF-T3 (#89) will build the Rust peer with the same tree, viewport, warmups, frame count, and scenarios. Until then, this lane is an explicit non-claim.
 
@@ -279,8 +271,6 @@ Trusted baseline (from `target/bench/tool-dispatch.json`, 2026-08-27, Xeon Gold
 | Rust CPU | 0.05 | 5.52% | pass |
 | TypeScript CPU | 0.102231 | 4.05% | pass |
 
-All four distributions pass the noise gate; per-implementation collection wall exceeds 1 s (10 samples × 10 000 calls each); order alternates per sample.
-
 Claim class: paired comparative. Wall ratio TS/Rust 0.77x (TypeScript's slice is faster in wall time); CPU ratio TS/Rust 2.04x (Rust uses half the CPU per call). Recorded as lane data — the Rust dispatch slice pays tokio task-spawn scheduling for the production parallel batch path, while upstream executes the batch on promises. Both implementations reject the shared invalid payload (`count: 999`) during argument validation with `update=0` and an error result per call; upstream additionally coerces mistyped primitives (TypeBox `Value.Convert`) where Rust rejects them — a recorded validation divergence, kept out of the timed payload.
 
 ### Lane 9: Session persistence/reopen (D8-blocked)
@@ -306,7 +296,7 @@ Hot/cold unit split:
 | Session/tree state reconstruction | Hot | Per-session; input-scaling |
 | Page-cache warm/cold | Cold | Cache-dependent; requires explicit warm/cold lanes |
 
-All persistence units are hot (>=5% of time during append/reopen). The cold-cache lane (page-cache miss on reopen) is a cold unit that requires explicit instrumentation.
+The cold-cache lane (page-cache miss on reopen) is a cold unit that requires explicit instrumentation.
 
 Claim class: D8-blocked. Composite evidence exists; no isolated timed boundary. PERF-T4 (#86) will add isolated session append and reopen timing lanes. Until then, this lane is an explicit non-claim.
 
@@ -378,8 +368,6 @@ Claim class: paired comparative, carefully named. The comparison supports "Rust
 
 ## Lanes with trusted baselines
 
-Four lanes (1, 2, 3, 11) have trusted baselines. Three are paired comparative timing lanes; one is an artifact-size comparison.
-
 | Lane | Rust median | TS median | TS/Rust ratio | Noise gate | Alternating | >=1 s wall |
 |---|---|---|---|---|---|---|
 | Startup `--version` (cold) | 40.07 ms | 540.42 ms | 13.49x | pass | yes | yes (TS) |
@@ -411,8 +399,6 @@ Four lanes (1, 2, 3, 11) have trusted baselines. Three are paired comparative ti
 | `pi-ai` | 2, 3, 8, 9, 10 | 2, 3 | none | 8, 9, 10 |
 | `pi-agent` | 3, 8 | 3 | none | 8 |
 
-All five crates have at least one paired comparative lane with a trusted baseline (lane 3 crosses all five crates). `pi-agent` has no standalone lane; its tool-dispatch path is measured through lane 8 (`pi` entry, `pi-agent` dispatch).
-
 ## Evidence provenance
 
 - `target/bench/performance-comparison.json`: check 9 artifact, generated 2026-08-26T14:22:29Z on this machine.

docs/PERF-R8-paired-baselines.md

@@ -22,19 +22,8 @@ PERF-T3 through PERF-T7. All five sibling tasks are now closed:
 | PERF-T7 | #91 | Lane 11 footprint: install accounting | CLOSED |
 
 This document records the trusted paired baselines on each newly symmetric
-lane, applies the same three-criterion trusted-baseline test from PERF-R2
-(noise gate < 20% relative spread, alternating order, >= 1 s collection wall
-for timing lanes), and updates the hot list.
-
-## Trusted-baseline criteria (from PERF-R2)
-
-1. **Noise gate**: median distribution with relative spread (stddev/median)
-   < 20%, enforced by `scripts/statistics.ts` `requireQuiet` at
-   `NOISE_RELATIVE_SPREAD_LIMIT = 0.2`.
-2. **Alternating order**: implementations alternate per sample index to
-   cancel sequential drift.
-3. **Collection wall**: per-implementation collection wall >= 1 s for timing
-   lanes (not applicable to artifact-size or allocation-only lanes).
+lane, applies the same three-criterion trusted-baseline test from PERF-R2,
+and updates the hot list.
 
 ## Lane 7: Layout/recomposition render churn (paired comparative)
 
@@ -56,10 +45,6 @@ Trusted baseline (wall ms/frame, 10 runs each):
 | TS static | 0.1120 | 18.45% | PASS |
 | TS editor | 0.2430 | 15.18% | PASS |
 
-All four distributions pass the noise gate. Collection wall per run is ~0.5 s
-(Rust) and ~0.7 s (TS); 10 runs give > 1 s total collection wall per
-implementation. Trusted baseline established.
-
 Allocation (KiB/frame):
 
 | Distribution | Median (KiB/frame) | Rel. spread | Notes |
@@ -155,18 +140,12 @@ Paired measurements (both implementations pass noise gate; blocked order, not al
 | append | warm | 1000 | 15.023 | 14.413 | 0.96x | 0.30 s / 0.29 s (< 1 s) |
 | append | warm | 5000 | 116.416 | 134.552 | 1.16x | 2.33 s / 2.69 s (>= 1 s) |
 
-Of these, only the two 5000-entry append cells meet the collection-wall
-criterion (>= 1 s per implementation). The 1000-entry warm cell passes the
-noise gate but collects < 1 s per implementation and uses blocked order, so
-it does not meet all three trusted-baseline criteria.
-
 Claim class: paired measurement with methodology gap (blocked order, not
 alternating). At 5000 entries, Rust append is 14-16% faster than TypeScript
 in both cold and warm cache. At 1000 warm entries, the two are within 4%
 (TS slightly faster). Peak RSS: Rust 2-16 MB vs TypeScript 52-105 MB across
 all cells — Rust uses ~7-25x less memory for session persistence.
 
-
 Hot list impact: **adds** JSONL append (hot, per-turn, >= 5% of time during
 append at 1000+ entries) and reopen (hot, per-session, >= 5% of time during
 reopen at 1000+ entries). **Removes** nothing — this lane was D8-blocked in
@@ -195,10 +174,6 @@ Trusted baseline (from artifact):
 | TS wall | 0.0177 | 9.14% | PASS |
 | TS CPU | 0.0999 | 4.71% | PASS |
 
-All four distributions pass the noise gate. Per-implementation collection
-wall exceeds 1 s (10 samples x 10 000 calls each). Trusted baseline
-established.
-
 Claim class: paired comparative. Wall ratio TS/Rust 0.75x (TypeScript's
 slice is faster in wall time — the Rust dispatch slice pays tokio task-spawn
 scheduling for the production parallel batch path). CPU ratio TS/Rust 2.04x
@@ -327,7 +302,7 @@ lanes and confirms one existing baseline. No units are removed.
 
 ### Lanes with trusted baselines after R8
 
-Seven lanes now have trusted baselines or paired measurements (up from five in R2's amended text — R2's summary tables listed four, but R2's Lane 8 section was amended to record a fifth):
+Seven lanes now have trusted baselines or paired measurements:
 
 | Lane | Rust median | TS median | TS/Rust ratio | Noise gate | Type |
 |------|-------------|-----------|---------------|------------|------|

docs/REL-R1-musl-toolchain-bakeoff.md

@@ -432,10 +432,6 @@ names. Primary-source verification performed 2026-08-27:
 | crates.io API | cargo-zigbuild 0.23.0 checksum | matches pin `68c7df45...`; not yanked |
 | GitHub runner-images README | `ubuntu-24.04-arm` | available (arm64) |
 
-All four aarch64 direct-download artifacts were downloaded and sha256-verified
-on the x86_64 host. Build and smoke execution is pending REL-T3 CI on
-`ubuntu-24.04-arm`.
-
 ### musl.cc probe (Candidate C rejection evidence)
 
 ```

docs/REL-R2-macos-signing.md

@@ -67,7 +67,6 @@ those facts outside this repository.
 ### Required credential set
 Confirm or provision every item below before enabling the signed channel.
 
-
 | Credential | Purpose | Who confirms or provisions | Notes |
 |---|---|---|---|
 | Apple Developer Program membership | Prerequisite for every Developer ID / notarization credential | Apple Developer account owner | Confirm active membership; enroll only if absent. |
@@ -175,11 +174,6 @@ repository administrator confirms or provisions the required Actions secrets.
 The required tools are already present on the `macos-15-intel` and `macos-15`
 runners at the dated image snapshot.
 
-This verdict is **non-gating by design**: the unsigned seven-target release definition is
-unaffected and remains the current ship contract. The darwin legs already build and package
-correctly unsigned; signing/notarization is a credential-gated follow-on, as the issue
-states.
-
 ## 7. Open items for the REL implementer (out of scope here)
 
 - Archive/container strategy for notarization: `notarytool` accepts ZIP, DMG, and PKG, but

docs/REL-R3-conpty-witness-prototype.md

@@ -324,11 +324,7 @@ the conhost-reap condition). The repo's shipped testkit already pins the
 conhost-specific pieces (`ConhostVtDec2026Fallback` profile, alt-buffer-aware
 snapshots) that make the assertions deterministic, and the windows-2025
 transcript leg already exercises the same portable-pty ConPTY driver against
-the fixture. The word *provisional* is load-bearing: this GO is derived from
-the exact vendored portable-pty 0.9.0 source, `microsoft/terminal@main`, and
-Microsoft Learn — it is not an executed windows-latest transcript. REL-T7
-converts it to a final GO by running the harness on windows-latest and
-requiring exit 0 with the artifacts of §2.
+the fixture. REL-T7 converts this to a final GO by running the harness on windows-latest and requiring exit 0 with the artifacts of §2.
 
 ### 5.2 Conditions attached to the GO (all implementable in REL-T7 wiring)