# Full graph assembly accepted and installed

2026-10-10 KST. [Workbench PR54](https://github.com/kimjooyoon/gooo-ecosystem-workbench/pull/54),
candidate `2211037ce4d6c5aaf48778253c943667589286c5`, was normally squashed with
the expected head into main `0d7d30a9e89576b4963e18b0967a71836849c234`.
Its sole parent is `7d64f52905e16d150baaa9c685c0ad5740a860a6`; its tree is exactly
the tested candidate tree. No reset, rebase or force push.

## Original CI and description

Original full [PR run37956868551](https://github.com/kimjooyoon/gooo-ecosystem-workbench/actions/runs/37956868551),
attempt1, exact candidate, SUCCESS. Job113909491338 has all38 steps SUCCESS.
The complete native race suite passed: root197.082s and CLI1.019s. Explicit
existing CLI examples and the other packages also passed. Separate original
push run37956858691 attempt1 also SUCCESS. Neither run was cancelled or restarted.

final-ci-run.json, final-ci-jobs.json and final-ci-artifacts.json retain the
actual original API responses. final-push-run.json and final-pr-rollup.json
confirm both checks completed before merge. pr-created.json, pr-before-merge.json
and pr-merged.json preserve the original description, compared unchanged before
and after acceptance. accepted-commit.txt, accepted-head.txt and accepted-tree.txt
record the normal accepted commit.

The successful raw job log was retained byte for byte:

- Plain bytes1,559,318; SHA256 `6d5942cc011c4ba65d5f9121e5acecec1d6be7e8a472da10424a4b515732a9d1`.
- final-ci-job.log.gz bytes244,448; SHA256 `f719eef67a78bce1ad3e3b22d6d73340a1de1ef8cada68e4ecdc780dcb764672`.
- Gzip has no timestamp. Decompression was compared with the retained original.

The first CLI fetch refused terminal escape sequences and returned no log bytes.
The successful fetch explicitly allowed those bytes. The compressed file above
is the unchanged successful response. The artifact metadata records
artifact11629085180, bytes13,919,462 and digest
`e49b7ac0482459b79eedab607cc0ec9cf54e15635f6c06c74a24a41226eafcdc`.
Its ZIP was not downloaded. Separate post-merge main run37958035452 attempt1 was
also confirmed terminal SUCCESS at exact accepted0d7d30a9 (postmerge-main-run.json).
The merge proof is the completed original candidate CI and identical accepted tree.

## Clean installation and actual CLI use

Clean accepted main was built with Go1.27.2 and GOWORK=off. Compiled code, embedded
Gooo recipes/model assets and module inputs matched compiled-inputs.sha256 from
the earlier direct dogfood. The executable was atomically installed and
byte-compared. Installed SHA256:
`1667c66e5cf0ff269ee2f591d3997c39345adccfa60028a44e2c06bdb438d7e3`.
installed-build.txt.gz retains actual Go build metadata after dropping only the
first executable-path line; accepted revision and vcs.modified=false are present.

The installed CLI ran these fresh routes:

- Two-record graph with existing own QAT model: two fresh initial body predictions,
  different context hashes, complete native saved replay with zero new calls.
  Caller expectation0/3 remains partial; this installation smoke has no final
  independent accuracy claim.
- Saved modeled graph continued with the model omitted: original failing rows
  prepared and consumed; one bounded fresh round, two record receipts each
  reporting model_calls0. Original replay verified. The partial budget1 result
  remains in installed-model-fixed-next.json.
- Called source-fill helper: original failure consumed, three rounds and six
  repeated candidate attempts, final separate activity outputs4/4. Native final
  evaluation reports zero fresh inference and four caller root tuples outside
  consumed construction inputs. Local fill holdout and training exposure are
  separate scopes.
- Typed-path graph: initial execution and saved replay verified; Gooo reports
  expand-joint-profile for its current whole-program construction gap.
- Input-only graph: original integer9007199254740993 retained, caller0/0
  unmeasured, Gooo requests expectations; saved replay zero new calls.

installed-smoke.log is the actual assertion output. Installed reports and
compressed native modeled assembly/replay, deterministic construction round and
final fill evaluation retain the observations. Help stdout/stderr are compressed
unchanged. Compiler source is clean `a15b413412ad7b465a0567fb1db4de5706a8c2a8`,
Go1.27.2 / decision SDK0.2.26. Public compiler release22/d3 remains unchanged and
does not contain body-plan.

## Earlier six-profile experiment and retention

The [accepted public experiment](https://github.com/kimjooyoon/gooo-ecosystem-workbench/tree/0d7d30a9e89576b4963e18b0967a71836849c234/publication/full-graph-assembly-20261010)
has source/caller rows, original compressed native JSON, TDD failures/corrections,
per-body model calls, budget16 failure and unscored input. Its dirty development
build identity remains explicit. Fixed/model repeated attempts127/111 and final
separate named outputs2/2 apply to that finite example; a one-run timing
observation is not a controlled benchmark. Source-local cases, adaptive caller
cases, final separate cases and unknown training exposure are distinct.

The full local1011-file dogfood folder was archived to4.1MiB and extracted/hash
verified before removing the owned42MiB raw folder. Only the two owned temporary
workbench binaries were removed after installation. The installed executable,
source checkouts, compressed full local evidence and public observations remain.