docs(share): record the Dwarves go-live in Batch 9

[tieubao]

docs(share): record the Dwarves go-live in Batch 9

Proof of done

From docs/verification/one-host-per-tenant.md:

Proof of done: one host per tenant, batch 1

Date: 2026-10-01
Branch: docs/one-host-per-tenant
Spec: docs/specs/SPEC-008-one-host-per-tenant.md

Batch 1 is TASK-1a and TASK-1b (live spikes; commands and answers in docs/implementation-notes/one-host-per-tenant.md) and TASK-2a (per-link storage on the origin, rows 2, 4, 5).

Green run

Command: SHARE_TEST_PORT_BASE=28787 gtimeout 900 bash tests/share.sh
Exit:    0
Checks:  1112 ok, 0 FAIL, 0 SKIP (PASS), at 3ea7dfe
Section: === tenant: per-link storage on the origin (rows 2, 4, 5) ===, 35 checks
Tail:    ok    no real Keychain share item changed during the run
         PASS
Verdict: PASS

Command: node tests/worker.mjs
Exit:    0
Tail:    ok    r28 malformed record aud
         PASS
Verdict: PASS

shellcheck bin/share install.sh tests/share.sh tests/e2e.sh tests/e2e-r2.sh demo/render.sh mac/*.sh and /bin/bash -n bin/share are clean on the same tree.

Negative controls

Each patch was applied to bin/share in the worktree, the tenant section was run on its own (the section's lines from tests/share.sh under a minimal check harness), then git checkout -- bin/share restored the file and the same run went green (35 ok, 0 FAIL).

Patch	Run	Red rows
write the pointer after stage_publish (the pointer line moved below the publish line in cmd_add)	exit with fails=4	row 4 "held at the pointer PUT, pub/ does not exist yet" (expected 1 0, got 1 1); row 5 "no pub/" (got 1 0 0 0, twice)
let a pointer PUT failure warn instead of die (r2_pointer_put's catch-all prints and returns 0)	exit with fails=4	row 5 "exits 1 naming R2" (got 0 0, ungated and gated); row 5 "no pub/, no row, no app" (got 1 1 0 0, and 1 1 1 0 gated: the app was created)
restore (git checkout -- bin/share)	fails=0	none

Dry trace for the first control: the row 4 run holds the add at PUT m/c00001 through SHARE_R2_DRY_PAUSE, whose .paused marker tells the test it is inside the held call; with the patch, stage_publish has already moved the stage to pub/c00001, so the check reads 1 where it wants 0.

Live spikes

TASK-1a and TASK-1b ran against throwaway Cloudflare objects only (names share-e2e-spk584462 and share-e2e-spk13b60d). Every object was deleted and checked from outside: the API listed none of them on either account afterwards, the minted token's id answered 404, and each zone's authoritative server answered NXDOMAIN for both names.

Batch 2

Batch 2 is TASK-2b, TASK-3, and TASK-4, each committed and checked on its own.

TASK-2b: dispatch, reconcile, sweep (rows 3, 6, 26, 27)

Command: SHARE_TEST_PORT_BASE=28787 gtimeout 900 bash tests/share.sh
Exit:    0
Checks:  1151 ok, 0 FAIL, 0 SKIP (PASS), at 6df0318
Section: === tenant: storage dispatch, member refusals, reconcile, sweep (rows 3, 6, 26, 27) ===, 39 checks
Verdict: PASS

node tests/worker.mjs PASS, and shellcheck bin/share install.sh tests/share.sh tests/e2e.sh tests/e2e-r2.sh demo/render.sh mac/*.sh clean, on the same tree.

Negative controls: each patch went into a copy of tree 25dd9e9 (commit 6df0318), the tenant sections ran alone, then the unpatched copy ran green (0 FAIL).

Patch	Red rows
let a member rm a machine row (the refusal in cmd_rm_r2 becomes a DELETE m/<id>)	row 3 "member rm of a machine row is refused" (got 0 0), "no DELETE m/, the pointer stays" (got 1), "member refresh ... refused" (got 1 0 0: the pointer was gone, so the message changed), row 27's machine rm (no pointer left to delete); fails=4
let the sweep count a machine pointer as unreadable (r2_record_raw's machine branch raises)	row 26 origin and member sweeps (got 0 0 1: skipped, warning printed), the not-JSON leg (got 0 0 0: the warning named the pointer instead); fails=3

Dry trace for the first control: the member's rm e10001 reads the snapshot, finds e10001 in r2_machine, and with the patch logs DELETE m/e10001; the check reads one DELETE line in the member's r2-calls.log where it wants none.

TASK-3: Worker v3 and r2_healthz (rows 7 to 10, 29)

Command: SHARE_TEST_PORT_BASE=28787 gtimeout 900 bash tests/share.sh
Exit:    0
Checks:  1211 ok, 0 FAIL, 0 SKIP (PASS), at 2fcbdb1
Section: === tenant: the Worker is up while the origin is off (row 29) ===, 6 checks
Verdict: PASS

Command: node tests/worker.mjs
Exit:    0
Checks:  134 ok, 0 FAIL (PASS), at 2fcbdb1; the tenant rows 7 to 10 add 58
Verdict: PASS

shellcheck and /bin/bash -n bin/share clean on the same tree. One earlier full run on this tree failed 500-row index answers under 3s once while the machine was busy; the rerun took 0.84 s and passed.

Negative controls on a copy of tree a6ff281 (commit 2fcbdb1):

Patch	Runner	Red	Green (unpatched)
the Worker passes an expired cloud record through instead of 404	node tests/worker.mjs	r7 an expired cloud record is 404, never passed through (got 200 1), plus the WORKER_SHA self-check	PASS
the Worker passes a miss through with PASS empty	node tests/worker.mjs	r9 PASS empty: a miss is 404 with no fetch call (got 101 1: the stub's last answer came back) and the machine-pointer leg (got 404 1), plus the WORKER_SHA self-check	PASS
read the healthz status alone in r2_healthz (the header branch never fires)	full suite	row 29 tunnel-down: state (got 0 false), the gated add (got 1), the join (got 1 0)	row 29 green; the copy has no .git, so the two v0.5.1 CLI checks that run git show fail in both runs (and profiles --json wrote nothing under HOME once in the red run, a flake); the worktree run above is 0 FAIL

Dry trace for the third control: the dry Worker writes HTTP/2 503 with the pair and x-share-tunnel: 0; with the patch hz_code stays 503, so state reports ready: false, the gated add dies at its healthz check, and the join's three-in-a-row wait times out.

TASK-4: setup --r2 and --no-r2, no alias (rows 11, 12, 13, 32)

Command: SHARE_TEST_PORT_BASE=28787 gtimeout 900 bash tests/share.sh
Exit:    0
Checks:  1243 ok, 0 FAIL, 0 SKIP (PASS), at 19926b1
Section: === tenant: setup --r2 and --no-r2 on the origin (rows 11, 12, 13, 32) ===, 32 checks
Verdict: PASS

node tests/worker.mjs PASS, shellcheck and /bin/bash -n bin/share clean on the same tree. Every call ran against the dry seam (SHARE_R2_DRY=1, a .test hostname); nothing was created on a Cloudflare account.

Negative controls on copies of tree a2cd324 (commit 19926b1), the full suite each:

Patch	Red	Green (unpatched copy)
attach the route before the pointer backfill (a route POST above step 9)	row 11 call order (stops at ^API POST /zones/zone-dry/workers/routes$), the rerun (a second route POST), row 13 (a route left after --no-r2), the fresh-bucket order; fails=7 with the copy's two git show checks and one load flake	every tenant row green; the two git show checks and one load flake (cleanup: nocred1 row gone, an unrelated SPEC-004 row; both copies ran at once)
let SPEC-007's r2 setup take a tenant host with --force (the PASS refusal skipped when --force is set)	row 32 "a tenant Worker (PASS) is refused with --force, before any write" (got 0 0 1: the setup ran on and wrote); fails=3 with the two git show checks	row 32 green; only the two git show checks

The copies carry no .git, so the two v0.5.1 CLI checks that run git show fail in every copy run; the worktree run above is 0 FAIL.

Dry trace for the first control: the patched setup logs API POST /zones/zone-dry/workers/routes before the first PUT m/0a000, so the order check finds the route POST's first line above the pointer PUTs and stops at that pattern; the rerun POSTs a second route because the patched line runs on every setup.

Batch 3

Batch 3 is the member hits wording (DEC-012), TASK-5, TASK-6, and TASK-7a, each committed and checked on its own. Every call ran against the dry seams (SHARE_R2_DRY=1, SHARE_ACCESS_DRY=1, .test hostnames); nothing was created on a Cloudflare account.

Green runs

Each run: SHARE_TEST_PORT_BASE=28787 gtimeout 900 bash tests/share.sh in the worktree, then node tests/worker.mjs, shellcheck bin/share install.sh tests/share.sh tests/e2e.sh tests/e2e-r2.sh demo/render.sh mac/*.sh, and /bin/bash -n bin/share.

Task	Commit	tests/share.sh	Exit	tests/worker.mjs	shellcheck
member hits wording	fcd24bd	1246 ok, 0 FAIL, PASS	0	PASS	clean
TASK-5 (rows 14, 30)	d246452	1281 ok, 0 FAIL, PASS	0	PASS	clean
TASK-6 (rows 1, 15, 16, 17, 28)	1f9fb4b	1312 ok, 0 FAIL, PASS	0	PASS	clean
TASK-7a (row 18, --token-stdin)	32a502e	1340 ok, 0 FAIL, PASS	0	PASS	clean

Negative controls

Each patch went into a git clone --local of the named commit in a mktemp directory (so the git show rows run as in the worktree), and the full suite ran there with SHARE_TEST_PORT_BASE=40000, one run at a time. An unpatched clone of the same commit ran green (EXIT=0) for d246452, 1f9fb4b, and 32a502e.

Patch	Commit	Red rows	Failed
member hits prints the bare count again (the note line becomes :)	fcd24bd	row 13 "no visits on an id another publisher added" and row 3 "member hits of a machine id never prints a bare 0" (both got 0 hits, 0 visitors alone)	2
keep the alias destinations on the app after step 14 (alias_dests drop skipped)	d246452	row 14 "dropped last, after the 301", row 14 end state, the rerun, and every row 30 convergence (each app still lists f.example.test/<id>{,/*})	7
print --no-r2 as the step-15 rollback with an alias (tenant_die ignores the alias)	d246452	row 30 at each of steps 12, 13, 14, 15 (got 1 0 0 0 1: no list, roll back with: share ... printed)	4
read the bucket in state through api_token_cmd (the stored-token check skipped)	1f9fb4b	row 17 "an api_token_cmd source" (got 0 1 6 0: the sentinel command ran, the bucket rows listed, no cloud_error)	3
merge bucket rows through rows() (the origin's bucket read rebinds index, outside the prune's subshell)	1f9fb4b	row 28 "the Caddyfile and index.tsv hold exactly the local rows" (the expired local row stayed, a bucket row counted in the index), plus rows 6, 11, 13, 14, 15, 30	45
let import accept a symlink (the pre-scan takes an l mode and a -> line, the post-scan a link)	32a502e	row 18 "a symlink is refused" (got 0 1 1: published, a row written)	1

Two unrelated lines failed in the red runs and are not counted as caught: row 29 "a member joins" in the api_token_cmd run (the healthz wait under load; green in every other run), and the real-launchd guard in both TASK-6 red runs. That guard saw a real foundation.d.share job appear on this machine during the first run and leave during the second. The suite stubs launchctl and labels its own jobs share-selftest-<pid>, so the job came from outside the suite; the worktree runs and the green clone saw no change.

An earlier attempt ran two clones at once; the green one failed prune still removes the payload and a row 27 rm with exit 5 while the twin run trashed same-named files. Every control above ran alone.

Dry trace for the TASK-5 rollback control: with the patch, the step-15 die at ALIAS-301 calls die "...; roll back with: share --profile ten setup ten.example.test --no-r2", so the output holds no alias rollback for f.example.test line and no 6. ... step, and row 30's count reads 0 where it wants 1.

Tar measurement

TASK-7a's first step ran on bsdtar 3.5.3 and GNU tar 1.35; the table and the flags it fixed are in docs/implementation-notes/one-host-per-tenant.md. Nothing was extracted outside the scratch directories: both tars stripped the absolute member into the target and refused the .. member, and the probe paths outside it did not exist afterwards.

Rollback

Batch 3 changes only bin/share, its tests, and docs on the PR branch; no release, tap bump, or Cloudflare object exists for it. Rolling it back is reverting commits fcd24bd, d246452, 1f9fb4b, and 32a502e on the branch.

Batch 4

Batch 4 is TASK-7b (migrate, rows 19, 20, 21, 31), committed at f0dc1ac. No dry seam exists for migrate (it refuses an R2-on tenant), so its local coverage is a curl/security shim pair plus SHARE_MIGRATE_SSH pointing at a wrapper that joins its own argv with spaces and replays it through fish -c (or sh -c) on the same machine, exactly mirroring what a real ssh round trip does to the command line; nothing was created on a Cloudflare account.

Green run

Command: SHARE_TEST_PORT_BASE=30787 gtimeout 900 bash tests/share.sh
Exit:    1 (one pre-existing, unrelated flake; see below)
Checks:  1360 ok, 1 FAIL
Section: === migrate: moving a tenant's origin (rows 19, 20, 21, 31) ===, 20 checks, all ok
Tail:    ok    no real launchd or systemd share job appeared during the run
         ok    no real Keychain share item changed during the run

Command: node tests/worker.mjs
Exit:    0
Verdict: PASS

shellcheck bin/share install.sh tests/share.sh tests/e2e.sh tests/e2e-r2.sh tests/e2e-tenant.sh demo/render.sh mac/*.sh and /bin/bash -n bin/share are clean on the same tree (tests/e2e-migrate.sh does not exist yet; it is TASK-9b's deliverable).

The one failing line, profiles --json wrote nothing under HOME (a find -newer marker check, pre-existing code untouched by this batch), is unrelated to migrate: it passed on a first full run of the same tree and failed on a second run of the identical code, the same intermittent-under-load pattern the TASK-3 and TASK-6 batches above already recorded for other lines. A rerun confined to the migrate section alone (the 20 checks above) is green every time.

Negative controls

Each guarantee's patch ran as a one-off (a sed-patched copy of bin/share, the relevant migrate scenario run standalone against it, the stock file confirmed green immediately after). No patch was committed.

Guarantee	Patch	Red	Green (unpatched)
ids preserved	the remote import call for one id is rewritten to a different id before bin/share runs it	B's index never gains the original id (the renamed one lands instead)	row 19: B's index holds exactly A's own ids
source never deleted	migrate_retire's mv "$pub/$id" "$root/migrated/$id" replaced with rm -rf "$pub/$id"	migrated/<id> never exists after a run that otherwise completes	row 19: every moved id's tree lands in migrated/, byte for byte equal to B's copy
retire runs only after verify passes	the gated-link `access_probe_round ...		dieturned intoaccess_probe_round ...; true || die` (the die never fires)

Dry trace for the first control: cmd_import's own id argument comes from the sender's positional argv, so a wrapper script that rewrites $2 before execing the real binary is enough to desync the id client-side of any base64 or checksum check; the manifest digest still matches (it is a digest of the tar's own bytes, not the id), so nothing else catches it except the id itself showing up wrong in B's index, which is exactly what the "ids preserved" guarantee asserts directly.

Rollback

Batch 4 changes only bin/share, tests/share.sh, and docs on the PR branch; no release, tap bump, or Cloudflare object exists for it. Rolling it back is reverting commit f0dc1ac on the branch.

Batch 5

Batch 5 is the kit:battery review fix pass on this PR (five findings plus two items raised from the live rehearsal), commits a045013..2a0fb65. Each finding carries its own commit with a failing test shown red before the fix, then green after; the fast per-finding reruns below are scratch slices of the real tests/share.sh (same fixtures and helper functions, the unrelated middle of the file cut out) used purely to get a faster red/green loop during the fix; no scratch file was committed.

Findings and their red/green

#	Finding	Fix	Red (before)	Green (after)
1	migrate handed the target no --tunnel-name; a source on the default name collided, the target reused the source's own tunnel, and migrate_retire then deleted it out from under the new origin	always pass a deterministic --tunnel-name distinct from the source's own (-m suffix); migrate_retire also refuses the tunnel/DNS deletes if the target's own tunnel id is ever reported equal to this machine's	row 19m: target's tunnel_id equals the source's	row 19m: target's tunnel_id never equals the source's; rows 19/20/21/31 unaffected
2	tests/share.sh printed the same "PASS"/"N FAILED" summary whether the run finished or crashed partway	a reached_end flag set only at the true end; the EXIT trap reports ABORTED and exits 2 when it is missing; the summary prints the total check count	isolated trap test: a truncated run exits 0 with no ABORTED text	isolated trap test, then the embedded self-test: a truncated run exits 2 and prints ABORTED
3	cmd_setup's old_host guard only fires when old_host is set; import into a not_setup profile then setup <other-host> served its gated rows ungated	refuse when old_host is empty and the profile holds an access= row; migrate's own switch (--token-stdin) is exempt, since it never renames the hostname	setup on a not_setup profile with a gated row: no refusal, dies on the live Cloudflare call instead (with the nocurl stub, before any call assertion fails)	refused before any Cloudflare call, naming the gated row
4	the Worker's pass() treated any 502 as the whole machine offline, same as Cloudflare's 520-527 edge codes	drop the explicit 502 check; only 530, 520-527, and a 503 with no cf-cache-status mean offline	node tests/worker.mjs: a dead-port 502 rewritten to the 503 offline page	the 502 passes through unchanged, with or without cf-cache-status
5	implementation notes claimed the migrate preflight reads tunnel_name from the API by tunnel_id when the config key is missing; the code never did	cmd_migrate now reads it back from the API when the local key is absent, and dies naming the manual fix if neither is known; note corrected to say where this runs	row 31b: rollback line printed --tunnel-name --force (empty)	row 31b: names the tunnel read back by id, never empty
6	alias_dests drop answered HTTP 400 live (3/3 rehearsals), the same PUT shape that worked for add two calls earlier	the PUT body drops the legacy self_hosted_domains/domain mirror fields so the API recomputes them from the new destinations, instead of echoing back a stale value that no longer names a surviving destination once drop removes it	row 14 ("the fold exits 0") against a genuinely reverted alias_dests, cf_dry extended to reject the same self_hosted_domains/destinations inconsistency a real PUT does, fixture carrying a realistic self_hosted_domains: fold dies, destinations keep the alias's own entries. Confirmed again with an isolated jq trace of the exact body construction (same expressions, no harness): pre-fix body retains self_hosted_domains naming a URI absent from the new destinations	row 14/30 (40+ checks) all pass end to end; the same isolated jq trace shows the stripped body consistent
7	tests/e2e-tenant.sh had five script bugs found across three live rehearsals (a cold R2 token, a split TCP read in the echo origin, a missing-file check, two Access-app lookups assuming a folded app renames to the tenant), plus L8's etag_before reading through A's own profile right after --no-r2 removed its bucket=	applied the run1→run3 fix patch; L8 now reads the pointer's ETag by a direct S3 HEAD on the bucket name, independent of A's profile config	(live-only; no dry seam exercises the R2/Access live path)	bash -n tests/e2e-tenant.sh and shellcheck clean; the lead reruns R1/L8 live

Green run

Command: SHARE_TEST_PORT_BASE=32787 gtimeout 900 bash tests/share.sh
Exit:    0
Checks:  1238 checks, PASS
Tail:    ok    a run truncated before the end exits 2
         ok    a run truncated before the end reports ABORTED
         ok    no real launchd or systemd share job appeared during the run
         ok    no real Keychain share item changed during the run

A first run of the same tree, same command, showed one failing line unrelated to any of the seven items above (row 6: pub/<id> absent while a PROBE fail was logged, a sleep 0.05 filesystem-polling race against a background watcher, pre-existing and untouched by this batch): the same intermittent-under-load pattern Batch 4 already recorded for a different line. The rerun above is clean.

Command: node tests/worker.mjs
Exit:    0
Verdict: PASS

Command: swift test --package-path mac
Exit:    0
Verdict: Executed 176 tests, with 0 failures (0 unexpected)

shellcheck bin/share install.sh tests/share.sh tests/e2e.sh tests/e2e-r2.sh tests/e2e-tenant.sh demo/render.sh mac/*.sh and /bin/bash -n bin/share tests/share.sh tests/e2e-tenant.sh are clean on the same tree.

Rollback

Batch 5 changes bin/share, tests/share.sh, tests/worker.mjs, tests/e2e-tenant.sh, and docs/implementation-notes/one-host-per-tenant.md on the PR branch; no release, tap bump, or Cloudflare object exists for it. Rolling it back is reverting commits a045013..2a0fb65 on the branch.

Batch 6

Batch 6 is the live rerun of tests/e2e-tenant.sh on the fix-pass head (1440259), the fixes that rerun forced, the TASK-10 docs, and the negative controls the earlier batches had not covered. Final live result: 73/73, run log docs/verification/e2e-tenant-20261005T042854Z.log. TASK-9b (tests/e2e-migrate.sh, rehearsal R2) is not run here: it needs one command from Han on the Air.

Live e2e: five runs, one green

Every run used the PR worktree's bin/share (SHARE_BIN), zone d.foundation, an admin token from op://Toolkit/cf-api-token/credential, run-scoped 2h publisher tokens minted through op://Toolkit/cf-tokens-admin/credential, and SHARE_E2E_ACCESS_RULE=group:dwarves-ops. Hostnames were share-e2e-<6 hex> pairs chosen per run. The earlier logs are not committed (two of them print the account and zone ids in a rollback line).

Run	Head	Result	What failed	Cause and action
1	1440259	69/73	L3 WebSocket echo; R1 rollback x2; R1 "alias bucket gone"	R1's fold passed for the first time (alias 301, folded gated link 302 with kid == aud, second fold), so the live 400 is fixed. The rest: the WebSocket client used the local resolver, which answered "No route to host" for a name made seconds earlier (script bug, now DoH). setup --no-r2 exited 1 with no output (a real bug in alias_rollback, below). The cleanup deleted the alias bucket with an admin env token on a profile with no stored publisher token (script bug)
2	f6a6f41	71/73	R1 rollback x2	the script parsed r2-call's code=<n> etag=<etag> line as etag= alone, so the marker host restore went out with an empty If-Match and the marker kept its aliases (script bug)
3	dac8541	70/72	R1 standalone origin and gated link	Cloudflare answered HTTP 500 adding the Worker custom domain for the alias host; the gated R1 legs did not run. Transient; the die names the call
4	dac8541	72/73	L1 "gated local link 302s to Access with kid == aud": got 200	the first fetch after add returned reached an edge that did not yet enforce the app. Not reproduced in run 5; see the note below
5	dac8541	73/73 PASS	none	R1 (fold, rollback, second fold, cleanup) green end to end

R1 passes: the fold exits 0; https://<alias>/healthz answers 301 to the tenant; the folded gated link 302s on the tenant with kid == aud; the rollback list runs in its printed order and --no-r2 leaves a plain tunnel (no route); the second fold converges and the alias 301s again; the cleanup deletes the bucket, the Worker, and the tenant DNS.

Gate note from run 4: share's own check passed three rounds of kid == aud before the link was published, and a fetch through DoH one second later still answered 200 once. Access enforcement is per edge and eventual (ADR-0006 measured seconds to minutes). The e2e keeps the strict check, because "no ungated window after add returns" is the guarantee the leg tests. A lead decision, outside SPEC-008: whether add --access should probe through a second resolver before it publishes.

Cleanup proof, from outside the run

A snapshot script listed every resource class the e2e touches, before and after each run, from the Cloudflare API and launchctl. Names and truncated ids only. The baseline and every later snapshot were identical (diff empty) after runs 1, 2, 3, 4, and 5.

Class	Baseline (and after every run)
Workers named share-*	share-f-d-foundation
Worker custom domains, routes, DNS records matching share-e2e*	none
Tunnels named share-* (not deleted)	share-s-d-foundation
R2 buckets named share-*	share-dfoundation
Access apps named share ...	s.d.foundation, f.d.foundation (the two live Dwarves apps)
Minted user tokens (a-origin, b-member, d-fresh, c-admin)	none
launchd labels and LaunchAgents plists matching share	foundation.d.share.dfoundation, the Share Bar app job

The script's own leftovers sweep printed cleanup: nothing left for this run's names on all five runs. No process named echo.py or share-e2e remained. Nothing the run did not create was touched: the two live Access apps, share-f-d-foundation, share-s-d-foundation, and share-dfoundation are in the baseline and in every later snapshot.

Findings the rerun forced

#	Finding	Fix	Red	Green
1	alias_rollback assigned ptrs from a pipeline ending in grep; with no local rows and no machine records the grep matches nothing, and under set -e the run died with exit 1 and no output, so --no-r2 and every tenant_die printed nothing for a tenant with an alias and no local rows (e6b2a13)	{ grep ... || true; } in the pipeline	the new row 30 check, run against the unfixed bin/share in a clone: row 30: with no local rows and no pointers, --no-r2 still refuses and prints the list: expected '1 1 1 1', got '1 0 0 0', 1 FAIL of 1243	the same check on the fixed tree: 1243 checks, PASS
2	five tests/e2e-tenant.sh script bugs (743b229, dac8541): DoH for the WebSocket client; the R1 rollback removes aliases= from the profile config (step 4 of the printed list); the marker ETag parsed from code=<n> etag=<etag>; the alias liveness check reads /healthz (a standalone r2 Worker answers 404 at the site root by design); the cleanup's delete-prefix uses the SHARE_R2_TOKEN seam	script edits	runs 1 and 2	run 5, 73/73

Negative controls

The spec lists 20. Batches 1 to 5 recorded a red run for 14 of them (the expired-record, PASS-empty, pointer-order, pointer-failure, route-order, alias-destination, symlink, api_token_cmd, rows()-merge, SPEC-007 --force, sweep, member-rm, healthz-status, and --no-r2-rollback controls). Batch 6 adds 5 and records one whose spec form is not reachable, with a substitute. Each patch went into a git clone --local of dac8541 in a scratch directory and the suite (or node tests/worker.mjs for a Worker patch) ran there, one run at a time with its own SHARE_TEST_PORT_BASE. The green run for every row is the unpatched clone of the same commit.

Green (unpatched dac8541)	Result
SHARE_TEST_PORT_BASE=41100 gtimeout 900 bash tests/share.sh	exit 0, 1243 checks, PASS
node tests/worker.mjs	exit 0, PASS

Spec control	Patch	Runner	Red
the Worker serves R2 for a storage:"machine" record (row 7), tested in the inverse form	the machine-pointer branch in worker_js made unreachable (&& false), so a pointer no longer passes through	node tests/worker.mjs	r7 a machine record passes through (got 404, no fetch) and the r8 pass-through rows (13 FAIL, plus WORKER_SHA drift, which any Worker edit trips)
drop the encoded-separator check before the pass-through (SPEC-007 row 18 on the v3 source)	the %2f |%5c |%2e |low escape |// test replaced with if (false)	node tests/worker.mjs	r18 for ..%2F, %2f, ..%5C, %2E, //, and a low escape (all 404, expected 400), and r7 /x/..%2Fb1c2d3/f.txt is 400 in front of the tunnel (12 FAIL plus WORKER_SHA)
take a member's add --local as cloud (row 3)	the member --local die replaced with storage=cloud	suite	row 3: member add --local f is refused (got 0 0), 1 FAIL
run the full teardown (with Access deletes) on the old origin before moving rows aside (row 19)	cmd_rm of every moved id at the top of migrate_retire	suite	row 19 migrate exits 0, B's trees equal A's, the rows move to index.migrated, the live row stays, migrate still exits 0; row 21 rerun skips the moved id (9 FAIL, one of them the harness's idler check after the abort)
omit --tunnel-name from the printed migrate rollback (row 31)	--tunnel-name $old_tunnel_name dropped from the switch-failure rollback line	suite	row 20: a remote-setup failure ... printing the rollback with A's own tunnel-name and row 31b (3 FAIL, one the idler check)
write bucket= into a plain tunnel config at setup (row 1)	echo "bucket=leak" added to write_config	suite	no red: 1243 checks, PASS. Not reachable: row 1's compat run does a quick setup only, and no suite row runs a named tunnel setup through write_config. The live e2e covers it (L1 setup, L7 --no-r2)
replacement for the row above: --no-r2 keeps bucket= (row 13)	tenant_config stops dropping bucket=	suite	rerun: the config is the same, row 13: the config is the one before setup --r2, row 13: the next add writes no pointer (3 FAIL)

One gap is open: the write_config row-1 control above. Closing it needs a named-setup fixture with a Cloudflare stub, which no suite row has; it is the one place where only the live e2e proves "a tenant with R2 off writes no bucket=".

Docs check (TASK-10)

tests/share.sh now checks that every verb, flag, config key, and state key the tenant docs name is in docs/how-it-works.md and in bin/share (the four generic state keys are checked in the doc only, since storage, type, by, and r2 are too common to grep in the source; rows 15 and 1 pin them in state output). It has three negative controls inside the suite: a doc missing --storage-default, a bin/share missing --remote-bin, and a doc missing the storage state key each fail the check. All pass in the green run above.

Rollback

Batch 6 changes bin/share (one line in alias_rollback), tests/share.sh, tests/e2e-tenant.sh, README, docs/how-it-works.md, docs/setup.md, ADR-0008, the spec (DEC-013, the 502 rows, Worker version 4), the implementation notes, and this record, on the PR branch. No release, tap bump, or standing Cloudflare object exists for it. Rolling it back is reverting commits e6b2a13..dac8541 and the docs commit that follows.

Batch 7

Batch 7 is the lead's three safety decisions and the CI fix, on top of Batch 6.

#	Change	Red (unfixed 476f4f8, new tests)	Green
1	migrate_retire refuses the DNS and tunnel delete when migrate-tunnel-id returns nothing, and names teardown --yes for later	row 19e: no DNS record or tunnel is deleted got 0 2; the refusal text absent	row 19e passes: migrate exits 0, rows moved aside, no DELETE of a DNS record or tunnel
2	the custom-domain add (standalone r2 setup and the alias bind) retries a Cloudflare 5xx, three tries with doubling backoff, and dies loud after (after up to 3 tries); the dry seam gains SHARE_R2_DRY_FAIL_TIMES	four checks: two 500s then success got 1 1, a persistent 500 got one PUT, not three	both paths succeed after two stubbed 500s (three PUTs) and die after three on a persistent 500; the alias die still prints the rollback list
3	add --access polls the published link until it answers 302 or 403 (SHARE_ACCESS_POST_WAIT, 30 s) before printing it; a link still answering anything else exits 1 with the status, the edge risk, and share rm <id>; the link stays published	four checks: the link printed with no probes, a 200 never failed the add	200,200,302 prints after three probes; 403 counts; a persistent 200 exits 1 with no link on stdout, and the warning names the rm command

Red run: 1256 checks, 11 FAIL (the 10 above plus one cascade from leftover gated rows, fixed in the test). Green runs on the final tree: 1256 checks, PASS, one earlier run showed a single row 29 timing failure (SHARE_R2_WAIT=1 join under load) that did not repeat. node tests/worker.mjs: PASS. Live: tests/e2e-tenant.sh on 1a76f2e (the first commit with the real link check against Cloudflare), 73/73, log docs/verification/e2e-tenant-20261005T062235Z.log; the before and after snapshots of every resource class are identical.

CI

The CI lint job failed on every push since the tenant work began, and the test jobs have failed since the migrate rows landed. Three causes, all fixed:

Job	Cause	Fix
lint	CI's shellcheck reports SC2015 twice in bin/share and SC2002 once in tests/share.sh; the local shellcheck 0.11 is quieter	restructured the three lines, same behavior; a later SC2001 in a new debug line fixed too
test (ubuntu, macos)	the migrate rows run setup on the target, which dies with missing: caddy cloudflared on a runner with no cloudflared; every dev machine has one	the suite appends stand-in caddy and cloudflared last on PATH, so a real binary still wins. Reproduced locally with a PATH that hides cloudflared and fish: red before, 1256 checks, PASS after

The macOS runner also showed one tty: the prompt appeared and the token did not echo failure on 1440259 that did not recur; it is the known flaky row.

Batch 8

Batch 8 is TASK-9b: tests/e2e-migrate.sh and the live rehearsal R2 of the personal move, on throwaway names. Final result: 38/38, run log docs/verification/e2e-migrate-20261005T093825Z.log. The second machine was the real Air, driven from the Mini with mini-run --host air, migrating over real ssh to the Mini.

Leg	What it ran	Result
M0	origin A (a snapshot, a folder, a gated snapshot, a live server), migrate to B on the same machine with a remote setup that does the real switch and then reports failure; the printed rollback run on A	exit 1 with the rollback line naming A's own tunnel; the DNS record moved to B's tunnel; the printed rollback exits 0 and the record points at A's own tunnel again; every link answers, the gated one with kid == aud
M1	the same name migrated again, B already holding the copies	exit 0; the copies are skipped; the live link is reported as not moved; every snapshot link answers at the same URL, the gated one keeps its 302 and AUD; the DNS record points at B's tunnel; A's tunnel is deleted; A's rows are in index.migrated and its trees under migrated/
M2	a second throwaway name set up on the Air, five links published, migrate --to the Mini over ssh	the same assertions as M1, from the real Air to the Mini

Measured gap: probes every 0.5 s through DoH on a snapshot link logged no non-200 answer across the switch on either run (37 probes in M1; the M2 gap logger likewise). The spec expected up to about a minute of 530; the DNS record flips and the new tunnel connects inside a probe interval at this scale, so the number to plan with is "under a few seconds", not a minute.

Findings the rehearsal forced

#	Finding	Fix	Red	Green
1	migrate archived only <name>, but a bare-file snapshot keeps its generated index.html beside the file at the id root. The receiver counted one file where the sender's manifest counted two and refused the share as a truncated copy: every single-file link would have failed the personal move (9ae6350)	the archive is the whole pub/<id> tree; import checks that <name> is in it	the suite fixture gains the generated index.html it lacked: 17 FAIL in rows 19 to 31	1256 checks, PASS
2	after a failed switch the target profile is serving this hostname, so the rerun's preflight died with port for this profile is already in use; the printed rollback led to a dead end (f37e816)	the preflight accepts a target already serving the same hostname	new row 19s: got 1 1	1257 checks, PASS
3	script bugs in tests/e2e-migrate.sh itself: profile names must be lowercase; the Air-side run needs the real HOME for ssh while share keeps a throwaway one; a new name needs a moment to resolve	script edits	runs 1 to 5	run 6, 38/38

Both product bugs live only on the real-account path: no fixture in the suite had a generated index beside a bare file or a target that was already serving.

Cleanup proof

Snapshots of every Cloudflare resource class (Workers, domains, routes, DNS, tunnels, buckets, Access apps, minted tokens) and of launchctl and LaunchAgents on the Mini were identical before and after the runs. On the Air, the count of foundation.d.share launchd jobs and the LaunchAgents listing were the same before and after (its real s.han.ws service was never touched: the legs used profiles mga and mgb under throwaway HOMEs). The script's own sweep printed cleanup: nothing left for this run's names on every run. One leftover surfaced: the M2 target's data root under the real HOME (~/share/profiles/mgb), because the script's cleanup only looked for the profile config. The script now moves that root aside too.

Rollback

Batch 8 adds tests/e2e-migrate.sh and changes bin/share (migrate_tar, the import root check, the migrate preflight) and tests/share.sh. No release, tap bump, or standing Cloudflare object exists for it.

Batch 9: live go-live, Dwarves tenant (2026-10-06)

Han typed the go in the operator session. Snapshots: tests/prod-snapshot.sh before D1 and before D3, kept outside git. The pre-D1 and pre-D3 snapshots differ only in share 0.8.0 vs share 0.9.0.

Step	Action	Check	Result
D1	merge #45 (a625255), bin/release --yes v0.9.0, notarized Share Bar 0.9.0, tap PRs homebrew-tools #28 and #29 merged, brew upgrade share on the Mini	brew list --versions share; status and live links unchanged	share 0.9.0; healthz 200 on both hosts; a68960 and ba6377 302
D3	setup s.d.foundation --r2 --bucket share-dfoundation --alias f.d.foundation (rc 0)	f.d.foundation/ba6377 301 to s.d.foundation, which 302s to Access with the snapshot AUD; a68960 302; ls lists a68960 machine and ba6377 cloud	pass
D2	minted the origin publisher token (180 days, expires 2027-04-04), stored in the Keychain from the Orca GUI session; removed the profile's api_token_cmd line that pointed at the admin token (config backed up first)	api-token --check	token found in the keychain, all lines ok, rc 0
D4	`share --pro

[cut at 40000 characters; the full proof is in https://github.com/dwarvesf/share/blob/973f673b313dbfd323c1817d02ebd4ade1fd3560/docs/verification/one-host-per-tenant.md]