Baffling sequential scan plan when index scan would be best

I always thought I would not be the kind of person who writes to this
list asking why the planner is using a sequential scan. I always looked
upon such people as newcomers who would eventually learn the mysterious
wonders of the Pg query execution planner.

But really, this plan is bizarre! Why is it scanning sequentially for
ONE tuple as selected by the primary key? I even increased stats to
1000 and disable seq_scan, but it still insists it cannot do an index
scan.

skunk=# \d items;
Table "items"
Column | Type | Modifiers
------------+----------+-----------
item | bigint | not null
[...]
Indexes:
"items_pkey" primary key, btree (item)

skunk=# analyze verbose items;
INFO: analyzing "items"
INFO: "items": 80372 pages, 300000 rows sampled, 2660996 estimated total rows
ANALYZE

skunk=# explain analyze select * from items where item = 2143888;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------
Seq Scan on items (cost=100000000.00..100113634.45 rows=1 width=115) (actual time=4034.564..8859.082 rows=1 loops=1)
Filter: (item = 2143888)
Total runtime: 8859.160 ms
(3 rows)

enable_hashagg | on
enable_hashjoin | on
enable_indexscan | on
enable_mergejoin | on
enable_nestloop | on
enable_seqscan | off <===
enable_sort | on
enable_tidscan | on

random_page_cost | 1
cpu_index_tuple_cost | 0.001
cpu_operator_cost | 0.0025
cpu_tuple_cost | 0.01

What's even more baffling is the planner will use index scan for any
other indexed column, including columns for which the index is not
particularly selective, like item category or date:

skunk=# set enable_seqscan=on;
SET

skunk=# explain select * from items where category = 245;
QUERY PLAN
--------------------------------------------------------------------------------------
Index Scan using items_cat_idx on items (cost=0.00..69887.77 rows=125795 width=115)
Index Cond: (category = 245)
(2 rows)

skunk=# explain select * from items where startdate = '2005-03-01';
QUERY PLAN
--------------------------------------------------------------------------------------
Index Scan using items_start_ids on items (cost=0.00..1948.53 rows=30283 width=115)
Index Cond: (startdate = '2005-03-01'::date)
(2 rows)

So it seems that an index scan returning a half-million tuples is OK,
but an index scan returning a single tuple is right out. What?

-Confused in California