## Evidence

**Additional git-count differences need maintainer review:** at the five-minute cache recurrence, total descendant git subprocesses are **10→40**, including retained root/fork **4→40**. Startup is **11→10**. D also has setup/control differences (deletion creation 20→18, root unarchive 13→14, root unsettle 18→19). These observations do not establish an overall git/CPU/disconnect reduction. Each bundle inventories all other count differences, including setup and unmatched provider-boundary requests.

Captured **15/19 matrix rows: all required R01–R15**, in six paired A–F runtime bundles. Optional S16–S19 and optional variants remain explicitly unverified below. Before [37fbefd](https://github.com/pingdotgg/t3code/commit/37fbefd147197fab9d88f9c120560c3f21b52cfc); after [45c1162](https://github.com/pingdotgg/t3code/commit/45c1162b91872e5e7a7ddd7843b24364caca0f4d). Source checkouts stayed clean; no code, PR body, or forge state was changed.

| Row | Fixture / mode | Before / after recordings and count tables |
|---|---|---|
| R01 | Empty / branchless | [Bundle A](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/76dff0510d7bf0ad/sheet-A.md) |
| R02 | Root + fork + three worktree delegates | [Bundle A](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/76dff0510d7bf0ad/sheet-A.md) |
| R03 | Shared app-owned delegate | [Bundle B](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/84f3f5104c834ced/sheet-B.md) |
| R04 | Shared provider-native child | [Bundle B](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/84f3f5104c834ced/sheet-B.md) |
| R05 | Direct fork of child | [Bundle A](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/76dff0510d7bf0ad/sheet-A.md) |
| R06 | Settled missing automatic reference | [Bundle C](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b0138927ce705814/sheet-C.md) |
| R07 | Settled saved automatic reference | [Bundle C](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b0138927ce705814/sheet-C.md) |
| R08 | Settled backfill exhaustion + recovery; disclosed git fault harness | [Bundle F](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/666b62050a35b2bf/sheet-F.md) |
| R09 | Archived / unarchived / deleted control | [Bundle C](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b0138927ce705814/sheet-C.md), [Bundle D](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4f457c9a5632c89/sheet-D.md) |
| R10 | Genuine created + metadata + unarchive child selectors | [Bundle D](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4f457c9a5632c89/sheet-D.md) |
| R11 | Unsettle + genuine checkpoint + settled metadata | [Bundle D](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4f457c9a5632c89/sheet-D.md) |
| R12 | Genuine completed child; remaining status/event selector harness | [Bundle D](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4f457c9a5632c89/sheet-D.md) |
| R13 | Headless policy gate + targeted child event | [Bundle E](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d032a8c99d56d907/sheet-E.md) |
| R14 | Existing PR discovery + persistence revalidation | [Bundle D](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4f457c9a5632c89/sheet-D.md) |
| R15 | No sweep-only child discovery, then child-event persistence | [Bundle D](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4f457c9a5632c89/sheet-D.md) |

| Core comparison | Before | After |
|---|---|---|
| Composite startup and each of six eligible minutes | 5 lookups | 2 lookups |
| Shared delegate, parent active | 1 | 1 |
| Shared delegate, parent settled after startup; local and native separately | 1 | 0 |
| Settled missing-reference startup; following minute | 3; 0 | 2; 0 |
| Each eligible child-specific selector | 1 | 1 |
| Headless startup; policy-gated minute; child event | 2; no request; 1 | 1; no request; 1 |
| Failed settled backfill attempts, root / child | 5 / 5 | 5 / 0 |
| Existing PR startup discovery+validation, root / child | 2 / 2 | 2 / 0 |
| First child event after sweep, discovery+validation total | 1 | 2 |

**Environment.** fs-18, Linux x86_64 6.1.102, Node v26.12.1; OpenCode 1.18.18 / `anthropic/claude-haiku-5-5`. Fresh independent homes: before `/tmp/lead-26983-r1/evidence-capture/20261029T150359-before-measured/home`, port 19684; after `/tmp/lead-26983-r1/evidence-capture/20261029T151950-after-final/home`, port 19685. `T3CODE_DEV_SEED=0` on every launch. Each recording identifies every activation’s actual server PID, captured spawn PID, SHA, clean status, home, port, observer hash and UTC times. Real one-minute scheduler and UTC clock; no fake/frozen clock. A scripted authenticated WebSocket client renews visible/focused activity every 15 seconds only in connected phases. No browser/UI surface applies.

**Capture method.** Completed `branchPullRequest` spans count through ancestry under `ThreadPullRequestServiceV2.synchronize`, separating discovery and persistence revalidation. Git counts cover descendant `runGitCommand` and git `processRunner.runProcessCore` executions, with cwd/branch, args, status and errors retained. Seed/provider/unrelated work is outside sweep totals. External observation adds labels without editing source; exact observer versions are supplied. F changes the executable only for disclosed deterministic `git remote` faults; R12 replays labelled internal selector events through the real service then drains the real worker. Genuine app-owned and native child creation, public RPC fork/settle/archive/etc., full lineage/origin and persisted automatic references are recorded.

**What else changed.** A: startup and expired-cache git counts above; warm minutes remain zero. B: measured minute git counts remain zero; native parent setup differs 0 L/1 G→1 L/2 G before fixture restart. C: missing-reference startup git 60→40; saved/archived outcomes remain zero. D: targeted lookup counts remain one; setup and replay git variations are fully tabulated, and automatic-reference validation occurs only when persistence is needed. E: first targeted headless child event is one lookup on both sides but 0→20 git subprocesses after excluded startup cache warming; setup root unsettle 8→15. F: root failures/retries stay present, child failures disappear; setup root unsettle 8→13. IDs, timestamps, PIDs, scratch paths and provider wording differ. Full non-equal request inventory: [count-differences.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ac35fe3a11109701/count-differences.json).

**Unverified runtime rows/variants.** S16 guards/cwd fallbacks; S17 replacement/summary branches; S18 races/errors; S19 private pending-membership transitions: optional supplemental harness rows in the supplied matrix, outside minimum A–F; not executed and no new runtime claim made. R05 linked-environment destination variant was optional and not run (single isolated environment used). R06 independent field-only settlement variants were optional and not injected (public settlement sets both fields). R15 minute-only external-PR creation variant was optional and not run (startup origin-transition variant captured; no forge writes authorized). Native completion is not claimed to produce a child terminal-run/checkpoint event.

**Rebuild and raw evidence.** [runtime-evidence.tar.gz](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e2ef042f050fc63c/runtime-evidence.tar.gz) contains sanitized original service traces/logs, ordered transcripts, full synthetic snapshots/event history, fault/replay inputs, all scripts (including prototypes), hashes, normalized fixtures, counter outputs and 69 capture consistency checks (69 passed). These checks inspect the recordings; they do not replace them. [reproduce.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1aaf0741be193d03/reproduce.md); [seed-all.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/20bdd79dd965e8b2/seed-all.py); [driver.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/cc6cf2896b85cc82/driver.py); [phases.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/10fd38ba5b43db57/phases.py); [counter.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/be83de27ac830a95/counter.py); [observer.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/57c6431d875b3d57/observer.mjs); [fault-phase.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/75c02f16f8d9bb59/fault-phase.py); [fault-git.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/15fcd2aabedcc803/fault-git.py); [known-pr.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f830341b046f349f/known-pr.py); [replay-selectors.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8c1070e72b974b51/replay-selectors.py). The fixture has a GitHub origin; existing public PR data was read only. No credentials, live database or real user data are uploaded. Earlier after pilots are segregated as excluded preflight.

Model/harness: GPT-6.1-sol through Codex; runtime fixture agents OpenCode / Claude Haiku 5.5.