Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/mips/common/util/gen-sjis-encode-table.py`:
- Around line 17-19: The if, elif, and else statements on lines 17-19 contain
multiple statements on a single line (condition and assignment together),
violating Ruff E701/E702 rules. Split each statement by moving the assignment of
base and lead_hi to a new line below the if/elif/else condition, so each line
contains only one statement. Apply the same fix to line 46 which has a similar
multi-statement violation.

In `@src/mips/psyqo/sio0-bus.hh`:
- Around line 46-47: The documentation comment at lines 46-47 is misleading
about how the re-entrant count works. The current wording implies that the
low-level driver function irqTransferBlocking() also acquires and releases the
lock, but it never actually calls acquire() or release(). Only the filesystem
layer (MemoryCardFileSystem) holds the lock across multiple sector operations.
Update the documentation to clarify that the re-entrant count exists because
filesystem operations hold the lock across all their sector transfers, not
because the blocking transfer function itself participates in arbitration. Make
it clear that only the filesystem operation caller (not the driver) acquires and
holds the bus.

In `@src/mips/psyqo/src/memory-card-filesystem.cpp`:
- Around line 229-238: The Phase::WriteDir case commits dirty directory entries
by index order, but this creates a safety issue with the multi-entry chain logic
built elsewhere (around lines 347-377) which requires the first entry of a chain
to be written last. Currently, if power loss occurs after writing the first
entry in a chain but before all linked entries are written, an incomplete file
chain can be exposed on the card. Modify the WriteDir phase to commit directory
entries in a specific order where the first entry of any chain is written LAST,
after all its linked entries are already committed. This ensures the "first
entry last" rule is applied consistently throughout the write and overwrite
paths, preventing partially committed files from being visible on interruption.
- Around line 444-455: The block chain traversal loop exits when the block value
goes out of range but still proceeds to the ReadTitle phase as if it completed
successfully. Add validation after the while loop to check if the chain
terminated cleanly by verifying that either the 0xffff terminator was
encountered or the chain length reached its maximum. If the loop exited due to
block going out of bounds without a clean termination, set m_result to
Error::BadData and return StepResult::Done instead of proceeding to the
ReadTitle phase.
- Around line 128-133: The totalBytes calculation at line 128 uses uint32_t
which can overflow when m_dataLen is large, allowing the FileTooLarge check to
be bypassed silently. Before computing totalBytes, check if adding
m_headerFrames multiplied by MemoryCard::c_sectorSize and m_dataLen would
overflow a uint32_t, or alternatively compute totalBytes using a larger type
like uint64_t and verify the result does not exceed the uint32_t maximum before
proceeding with the block-count calculation and the existing bounds check.

---

Nitpick comments:
In `@src/mips/psyqo/src/memory-card.cpp`:
- Around line 718-758: The async sector transfer methods `readSector` and
`writeSector` initiate async operations via `startAsyncAttempt` without
acquiring the SIO0Bus lock, relying entirely on the caller to hold it and
creating a footgun for direct callers. To fix this, acquire SIO0Bus at the
beginning of `startAsyncAttempt` method and release it in the
`onAsyncAttemptDone` method that completes the async operation. Since SIO0Bus
uses a re-entrant counter, nested acquisition from the filesystem's outer lock
will be safely tolerated.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@src/mips/psyqo/src/memory-card-filesystem.cpp`:
- Around line 373-399: The nameLen variable is computed by scanning m_name
without bounds checking, and when copied to entry + c_offName using
__builtin_memcpy, any nameLen exceeding the field width of 20 bytes will
overflow into adjacent metadata fields. Additionally, if m_name is unterminated,
the while loop will read out of bounds. After computing nameLen (and before the
memcpy at line 391), clamp nameLen to c_maxNameLength to ensure the copied name
does not overflow the field width, matching the validation pattern used
elsewhere in this file.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@src/mips/common/util/sjis-encode.h`:
- Around line 38-85: The encode/emit loop is duplicated between utf8ToSjis and
utf8ToSjisTitle, with only the title-specific fullwidth codepoint remap
differing. Refactor the shared UTF-8 decode, unicodeToSjis lookup, fallback, and
bounded write logic into a common helper, and let utf8ToSjisTitle supply the
remap behavior while utf8ToSjis uses the default path. Keep the existing public
function names and preserve their current output semantics.