Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@docs/migration-prep/claude-code-chat-transcript-93-29-26.md`:
- Line 7: CI is failing due to codespell-detected typos in the transcript line
containing "▘▘ ▝▝                       │ No recent activit│"; update that line
to remove the stray box-drawing/garbage characters and correct "activit" to
"activity", and apply the same fix to the other affected transcript occurrences
(files/locations referenced as 88-88, 245-245, 356-356, 364-364) so the text
passes codespell linting.
- Around line 8-10: The file contains a direct personal email
"bbrenner2217@gmail.com" in the transcript header; remove or redact this address
by replacing it with a masked form (e.g., "bbrenner2217@***" or "[redacted]") or
delete the email line entirely so no personal contact info remains; locate the
header text containing "Sonnet 4.6 · Claude Pro" and the exact string
"bbrenner2217@gmail.com" and update it accordingly before merging.

In `@docs/migration-prep/clean_profile`:
- Around line 24-29: The script prepends $HOME/.local/bin twice and may add it
when the directory doesn't exist: remove the unconditional line that prepends
$HOME/.local/bin (the final export PATH="$HOME/.local/bin:$PATH") and either
leave the conditional block that prepends when [ -d "$HOME/.local/bin" ] or
replace that final line with a plain export PATH to export the variable without
modifying its contents; locate references to PATH and $HOME/.local/bin in the
same snippet to apply the change.

In `@docs/migration-prep/now.md`:
- Line 2: The branch label in the header "## 09:03 | master" may be stale;
update that label to match the repository's current default branch (e.g.,
replace "master" with "main") or make the header dynamic/neutral (e.g., "default
branch") so handoff steps aren't misleading—locate and change the header string
"## 09:03 | master" in docs/migration-prep/now.md accordingly.

In `@docs/migration-prep/phases/bd_setup.sh`:
- Line 1: Set the executable bit for the script
docs/migration-prep/phases/bd_setup.sh (the bd_setup.sh file with the
#!/usr/bin/env bash shebang) in git metadata and commit that change so CI passes
the check-shebang-scripts-are-executable rule; ensure the file retains its
shebang and only the executable permission is changed before committing.

In `@docs/migration-prep/phases/PHASE_01_DEPENDENCY_BASELINE.md`:
- Line 107: The file PHASE_01_DEPENDENCY_BASELINE.md is missing a trailing
newline which triggers the pre-commit end-of-file-fixer; open that file and
ensure it ends with exactly one final newline character (no extra blank lines),
save, and re-run the linter or pre-commit to verify the EOF fix.

In `@docs/migration-prep/phases/PHASE_06_SERVER_AND_HTTP_DRIFT.md`:
- Around line 44-45: The documented destination for the network utilities is
wrong; update the migration doc to reference the actual location
`python/sglang/srt/utils/network.py` (where functions like try_bind_socket live)
or, if you prefer the documented layout, move the network utilities from
`python/sglang/srt/utils/network.py` into `python/sglang/srt/network.py` and
update all imports accordingly (search for usages of try_bind_socket and adjust
import paths to match the chosen location).

In `@docs/migration-prep/phases/PHASE_08_FULL_VALIDATION_AND_STRESS_TEST.md`:
- Around line 88-89: The document has a contradiction between the "10f |
Resilience / crash testing | **PASS** (4/5 minimum; SIGTERM hang is known Bug
`#16`)" gate and the later "all 12 phases pass" completion criterion; decide which
policy to enforce and make the text consistent: either update the 10f gate text
(the "Resilience / crash testing" / "10f" line) to require 5/5 pass or relax the
completion criterion that references "all 12 phases pass" to allow one tolerated
partial pass (4/5) citing Bug `#16`, and update the accompanying explanatory note
to explicitly state the accepted exception and link to Bug `#16` so the gate is
unambiguous.

In `@docs/migration-prep/pull_from_runpod.sh`:
- Line 1: The script pull_from_runpod.sh has a shebang but is not executable;
run chmod +x pull_from_runpod.sh to set the executable bit and then stage and
commit the mode change so the repository records the file as executable; this
will satisfy the check-shebang-scripts-are-executable CI check and prevent the
lint job from failing.

In `@docs/migration-prep/settings.json`:
- Line 13: The committed default in settings.json enables the dangerous-mode
bypass via the "skipDangerousModePermissionPrompt" key; remove or set
"skipDangerousModePermissionPrompt" to false in the tracked
docs/migration-prep/settings.json and instead document in BOOTSTRAP_FRESH_VM.md
that users may optionally add a local override to ~/.claude/settings.json to opt
into skipping the prompt for a single machine; ensure the repo-default file does
not disable the safety catch.

In `@docs/migration-prep/vultr_startup.sh`:
- Line 1: The script begins with a shebang ("#!/bin/bash") but lacks the
executable bit in git metadata; update the file's mode to set the executable bit
(make the script executable in git) and commit that change so the pipeline lint
can run against the script.
- Around line 10-12: Guard the directory changes in the startup script by
enabling fail-fast (e.g., set -euo pipefail) and explicitly check the result of
each critical command such as the initial cd /root, the git clone
https://github.com/KHAEntertainment/sglang-mamba.git, and cd sglang-mamba; if
any of those commands fail, emit a clear error message and exit non‑zero (for
example, replace bare "cd /root" and "cd sglang-mamba" with guarded checks like
cd /root || { echo "Failed to cd /root"; exit 1; } and verify git clone
succeeded before attempting to cd into sglang-mamba).

In `@python/sglang/srt/entrypoints/http_server.py`:
- Around line 2244-2253: The pre-bound socket creation (reserved_socket,
set_inheritable, bind) must be performed only for the single-worker / FD-handoff
path; update the code that creates reserved_socket to first check the mode
(e.g., inspect server_args.workers or other flag used to decide the
single-worker branch and whether uvicorn.run is invoked with
fd=reserved_socket.fileno()), and skip creating/binding the socket when running
in multi-worker mode where uvicorn.run is called with host/port (or when workers
> 1); keep the existing reserved_socket usage for the fd-handled launch (and
ensure it remains inheritable there) and avoid the bind/creation in the
multi-process branch to prevent double-bind and wasted socket resources (refer
to reserved_socket, server_args.workers, uvicorn.run, and
monkey_patch_uvicorn_multiprocessing to locate the logic).

In `@test/phases/results/phase-03-upstream-sync-2026-03-31.md`:
- Around line 147-152: The fenced code block containing the commit list in
phase-03-upstream-sync-2026-03-31.md is missing a fence language and triggers
MD040; add a language identifier (e.g., "text") to the opening triple-backtick
so the block reads ```text to satisfy the linter and keep the existing commit
lines unchanged, ensuring the block around the four commit lines (47e7f0743,
1d906a894, 4b74fc896, 040c0a43a) is updated accordingly.

In `@test/phases/results/phase-04-upstream-sync-2026-03-31.md`:
- Line 42: The summary line stating "Total: 17 upstream commits integrated, 14
conflict-free, 3 with minor conflicts." is incorrect; recalculate the aggregates
to match the per-row data (rows in the table showing 6 commits with conflicts
and 11 conflict-free). Update the summary text so it reads "Total: 17 upstream
commits integrated, 11 conflict-free, 6 with conflicts" (or equivalent
phrasing), ensuring the counts in the summary match the table rows referenced in
the file.

In `@test/phases/results/phase-07-upstream-sync-2026-03-31.md`:
- Around line 69-74: The two fenced code blocks in the markdown are missing
language tags; locate the block that begins with
"test/sglang/snapshot/test_mamba_snapshot.py                      11 passed, 1
skipped" and the block that begins with "sglang.srt.snapshot                   
OK" and change their opening fences from ``` to ```text (or ```bash if more
appropriate) so both blocks have a language identifier and the MD040 lint
warning is resolved.
- Line 125: The markdown table row currently lists __init__ without code
formatting which triggers markdownlint strong-style warnings; update the table
entry so the initializer name is wrapped in inline code backticks
(`__init__`)—locate the table row containing "mamba_snapshot, tier_manager,
conversation_tracker, snapshot_policy, snapshot_hooks, mamba_host_pool,
__init__" and replace the bare __init__ with `__init__`.

In `@test/phases/results/phase-08-upstream-sync-validation-2026-03-31.md`:
- Around line 104-113: The table's totals row currently has only two cells;
update the row containing "**Total: 6/6**" so it has three pipe-separated cells
to match the table schema (Test | Requests | Result): put the total label in the
first cell ("**Total: 6/6**"), add an appropriate Requests summary (either the
summed requests value or a placeholder like "-" ) in the second cell, and put
the overall result ("**PASS**") in the third cell so the table conforms to the
3-column layout and avoids MD056.

In `@test/registered/radix_cache/test_mamba_metadata.py`:
- Around line 6-113: The tests currently use unittest (TestMamba2Metadata,
unittest.main()) and inline MagicMock setup; migrate to pytest by converting
TestMamba2Metadata methods into top-level pytest test functions that use
fixtures, remove unittest.main(), and add
register_cuda_ci(est_time=<appropriate_seconds>,
suite="stage-b-test-small-1-gpu") near the top to register Phase 5; create
fixtures for forward_batch (used by prepare_mixed tests) and for shared metadata
creation (use or adapt _make_forward_metadata) and update tests that call
Mamba2Metadata.prepare_decode and Mamba2Metadata.prepare_mixed to accept those
fixtures while preserving assertions about
num_prefills/num_decodes/mamba_cache_indices and mixed_metadata flags.

In `@test/registered/radix_cache/test_mamba_pool_extended.py`:
- Around line 1-14: The file interleaves executable calls register_cuda_ci and
register_amd_ci between import statements which violates isort import grouping;
move all import statements (including unittest, torch, and the sglang.* imports)
to the top of the file, then place the two calls to register_cuda_ci and
register_amd_ci (with their est_time and suite args) after the import block so
imports are grouped together and execution follows imports; update around the
symbols register_cuda_ci and register_amd_ci and ensure no other executable code
appears before the imports.

---

Nitpick comments:
In `@docs/migration-prep/phases/bd_setup.sh`:
- Around line 18-32: The script currently creates a .beads symlink which can
dirty the repo; update the bd_setup.sh logic to ensure the symlink path is
ignored by git before creating it: detect/append the literal ".beads" entry to
.gitignore if missing (use BEADS_REAL and the existing ln -sf ".beads" step as
the insertion point), or alternatively ensure any existing ignore rule covers
the symlink (not just ".beads/*.db"); keep this change idempotent so repeated
runs don't duplicate entries.

In `@docs/migration-prep/phases/PHASE_05_SCHEDULER_IDLE_AND_POOL_FIXES.md`:
- Around line 65-68: Update the "Validation Criteria" section under the
"Validation Criteria" header so each referenced phase explicitly shows its
expected test counts (e.g., "Phase 2 (5/5)", "Phase 3 (16/16)"), matching the
format used in PHASE_04 docs; edit the list item that currently reads "Phases 2,
3, 4, 5, 6, 7, and 8 all **PASS**" to enumerate each phase with its expected
pass count and total count, and ensure the wording about "No `mamba_lock_ref`
violations or pool exhaustion errors in logs" remains unchanged.

In `@docs/migration-prep/phases/SYNC_MASTER_PLAN.md`:
- Around line 33-35: Update the fenced code block containing the dependency
chain "PHASE_01 → PHASE_02 → PHASE_03 → PHASE_04 → PHASE_05 → PHASE_06 →
PHASE_06B → PHASE_07 → PHASE_08" to include a language identifier (e.g., add
"text" after the opening backticks) so the block becomes a labeled fenced code
block for consistent Markdown rendering.

In `@test/phases/results/phase-08-upstream-sync-validation-2026-03-31.md`:
- Around line 22-143: The markdown tables (e.g., the tables under "Phase 0 —
Environment Verification", "Phase 1 — Stateless Inference Baseline", "Phase 2 —
MambaPool Unit Tests", etc.) need blank lines inserted before and after each
table block so each table is separated by an empty line from surrounding text;
update the document by adding a single blank line above the table start (the
line with the header row like "| Test Suite | Tests | Passed | Result |") and a
single blank line after the table end (the last row line) for every table to
satisfy MD058 and reduce lint noise.

In `@test/registered/radix_cache/test_mamba_radix_cache_server_integration.py`:
- Around line 98-123: The test currently builds full-history replay by
repeatedly appending to messages and calling self._chat(...) without a request
identity, so it verifies stateless replay rather than server-held continuity;
either rename the test to indicate "replay" behavior or change the scenario to
assert continuity by using a persistent conversation id (add an "rid" or similar
identity to the self._chat calls) or by performing a snapshot/restore flow and
asserting state carried across restore; update references around the messages
list and the self._chat(...) calls (and the test function name) so the test name
matches the asserted behavior.