perf(odf): narrow the stored ids, sort the payload tables, size the sheet arrays

[andiwand]

🤖 Generated with Claude Code

Takes the rest of the measured ladder in #762. odr-private/ods/efficiency-big-1.ods
(1.4 MB, a 297 MB content.xml) goes from 914 MB to 626 MB of peak footprint
through a full html translation, at the same 2.6 s. Nothing here changes what is
rendered.

Measured with phys_footprint - the metric jetsam and lmkd actually use - on
macOS arm64, cli/translate to a file.

	now	before
inflated content.xml, which the dom points into	284 MB	284 MB
pugixml dom	116 MB	116 MB
elements	101 MB	177 MB
payload tables	34 MB	130 MB
sheet index, and the slack it grew through	17 MB	68 MB
peak through the translation	626 MB	914 MB

The three commits

perf(odf): store the registry's ids narrow. An element is five ids, a type
and a node pointer, and nothing else: 56 bytes at 64 bits per id, 32 at 32 bits.
ElementIdentifier itself stays 64 bits - it is public, and it is not everywhere
an index: csv has no registry and packs a kind, a row and a column into its
bits, so narrowing the public type truncates a csv document to nothing (which is
how it was found - html.a_sheet_written_in_full_reports_no_cut fails and then
crashes on the end iterator). StoredId narrows inside the odf registry alone,
every boundary widens back, and create_element refuses to hand out an id it
could not store. 914 MB → 777 MB.

perf(odf): keep the per-element payloads in sorted arrays. A text's last
node, a table's column chain, a sheet's index, a sheet cell's position were
unordered_maps keyed by element id. Every one is written the moment its element
is created and ids increase, so the parser only ever appends in key order: a
binary search replaces a hash map's node and bucket slot. libc++ spends 32 bytes
on a node carrying 16 of payload; m_texts and m_sheet_cells are one entry per
non-empty cell, so the four tables were 130 MB against 34 MB of array. A deque,
not a vector, because create_*_element hands back a reference the parser fills
in while it keeps parsing. 777 MB → 677 MB.

perf(odf): size a sheet's arrays before filling them. The index grows to
one entry per row and per cell node at most - a repeat collapses onto one - so a
walk of those nodes gives the size up front, and buys the doubling: a vector that
reaches a million entries by doubling holds a third more than it needs and peaks
holding both halves while it copies. The count is an upper bound bounded by the
dom that is already resident, so a repeat bomb still registers its handful of
entries. 677 MB → 626 MB, and no slower - the copies it saves pay for the walk.
Shrinking afterwards instead measured 76 MB worse: the copy lands exactly on
the high-water mark, which is what the peak is.

Verification

• Full odr_test green: 1262 pass, the 6 pre-existing skips.
• The whole reference-output corpus is byte for byte identical, public and
  private - every change here is representation, not meaning.
• No public header changed, so no binding and no consumer of the public API sees
  anything. Only our own two tests include the internal registry header.

Not in here

The two items left in #762 are different in kind, and each has its own issue now:

• A non-empty spreadsheet cell builds three elements before any html limit applies #781, an element per non-empty cell. A cell with content still builds a
  cell, a paragraph and a text before any HtmlConfig limit applies - 3.16M
  elements for 1.05M cells here, ~150 MB with the entries they own. Empty and
  repeated cells already cost nothing. Making the children lazy (Sheet::Cell
  already holds the node) is worth ~85 MB.
• Stream a spreadsheet's content.xml instead of holding all of it: parser options #782, the 284 MB source buffer. pugixml has no streaming api of any kind,
  and load_buffer_inplace_own means the buffer is the dom's string storage -
  the 400 MB of buffer plus nodes only goes away with a pull parser and a model
  that owns its strings, read-only, alongside the dom path that edit and save
  need.

With those two filed, #762 is done once this merges.